Risk control training sample optimization method and device, and electronic equipment

Abstract

The application provides a risk control training sample optimization method and device and electronic equipment, and relates to the technical field of computers. The method comprises the following steps: extracting feature paths corresponding to at least one target leaf node in a target decision tree model respectively, and determining the feature paths as feature combination rules; the target decision tree model is obtained by training based on a historical risk control data training set; extracting target training samples meeting each feature combination rule from the historical risk control data training set respectively, and determining each target training sample as a candidate positive sample; for each combination rule group, determining a similarity score corresponding to each candidate positive sample in all candidate positive samples corresponding to the combination rule group based on a target Gaussian mixture model corresponding to the combination rule group; and optimizing each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combination rule group, so as to realize optimization of each candidate positive sample, improve sample quality and business rationality, and improve the performance of a risk control model.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus and electronic device for optimizing risk control training samples. Background Technology

[0002] In the field of risk control modeling, the problem of imbalanced sample distribution has long constrained the improvement of model performance. Because the actual occurrence rate of risk events such as fraudulent transactions and black market activities is relatively low, the number of positive samples (risk samples) in the training set is often far less than the number of negative samples (normal samples). This imbalance in data distribution can cause the model to be overly biased towards the larger number of negative samples, thereby reducing its ability to identify positive samples.

[0003] Currently, common sample augmentation methods are mainly divided into three categories: oversampling (such as Synthetic Minority Oversampling Technique (SMOTE)), undersampling, and cost-sensitive learning. However, these methods have significant limitations in risk control scenarios: oversampling techniques (such as SMOTE) generate synthetic samples through interpolation, which may lead to distortion of the feature space, generating unreasonable noise data and affecting the model's generalization ability; random undersampling, although it can balance the data distribution, will lose a lot of valuable information and reduce the model's robustness; cost-sensitive learning, although it can adjust the classification threshold, cannot fundamentally solve the problem of insufficient positive samples, and relies on manual adjustment of weights, resulting in poor flexibility. In addition, most existing methods rely on numerical feature distribution and ignore the interpretability of business rules, which may lead to generated samples that do not conform to the actual risk control logic.

[0004] Therefore, how to provide a sample optimization method that can both increase the number of positive samples and ensure sample quality and business rationality is an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a risk control training sample optimization method, apparatus, and electronic device to solve the problem of how to provide a sample optimization method that can both increase the number of positive samples and ensure sample quality and business rationality.

[0006] This invention provides a method for optimizing risk control training samples, comprising: The feature paths corresponding to at least one target leaf node in the target decision tree model are extracted respectively, and the feature paths are determined as a feature combination rule; wherein, the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set includes multiple training samples of at least one dimension and label data of each training sample; the dimension is user behavior features and / or user transaction features; Target training samples that satisfy each of the feature combination rules are extracted from the historical risk control data training set, and each of the target training samples is determined as a candidate positive sample; For each set of combined rules, based on the target Gaussian mixture model corresponding to the combined rule set, a similarity score is determined for each candidate positive sample among all candidate positive samples corresponding to the combined rule set; the target Gaussian mixture model is trained based on the first positive sample corresponding to the common feature of the combined rule set and the candidate positive sample; the combined rule set includes at least two feature combination rules. Based on the similarity score corresponding to each candidate positive sample in the combined rule group, each candidate positive sample is optimized.

[0007] According to a risk control training sample optimization method provided by the present invention, the step of extracting the feature path corresponding to at least one target leaf node in the target decision tree model includes: Traverse each leaf node of the target decision tree model to determine the sample set corresponding to each leaf node; each leaf node includes at least one positive sample; the positive sample is determined based on the label data of each training sample; the sample set includes the at least one positive sample. Based on the number of each of the positive samples in the sample set, at least one target leaf node is determined from each of the leaf nodes; Based on each of the target leaf nodes, the feature path corresponding to each target leaf node is determined.

[0008] According to a risk control training sample optimization method provided by the present invention, the method further includes: Based on any two feature combination rules among all feature combination rules, determine the number of common features between the feature sets involved in the two feature combination rules. Based on the number of common features, determine the feature overlap between any two feature combination rules; Based on the overlap of each feature, a similarity matrix is ​​constructed; Based on the similarity matrix, all feature combination rules are grouped to obtain at least one combination rule group; For each group of combined rules, extract the common features of all combined rules in the group.

[0009] According to a risk control training sample optimization method provided by the present invention, the step of determining the similarity score corresponding to each candidate positive sample among all candidate positive samples corresponding to the combined rule group based on the target Gaussian mixture model corresponding to the combined rule group includes: Each candidate positive sample in the group of combined rules is input into the target Gaussian mixture model to obtain the log-likelihood value of each candidate positive sample output by the target Gaussian mixture model. The log-likelihood values ​​of all candidate positive samples corresponding to each of the combined rule groups are normalized to obtain the similarity score of each candidate positive sample among all candidate positive samples corresponding to the combined rule groups.

[0010] According to a risk control training sample optimization method provided by the present invention, the optimization of each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combined rule group includes: Based on the similarity score corresponding to each candidate positive sample in the combined rule group, the target quality screening threshold is determined; Based on the target quality screening threshold, each of the candidate positive samples is optimized.

[0011] According to a risk control training sample optimization method provided by the present invention, the step of determining the target quality screening threshold based on the similarity score corresponding to each candidate positive sample in the combined rule group includes: Based on the second positive sample and the first negative sample in the validation set, and the similarity score corresponding to each candidate positive sample, multiple similarity score screening thresholds are tested within the target range. The target quality screening threshold is determined from each of the aforementioned similarity score screening thresholds.

[0012] According to a risk control training sample optimization method provided by the present invention, determining the target quality screening threshold from each of the similarity score screening thresholds includes: Based on each of the similarity score screening thresholds, the candidate positive samples corresponding to similarity scores greater than the similarity score screening thresholds are determined as the expanded training samples. The expanded training samples are used to train the initial risk control model to obtain the target risk control model. The second positive sample and the first negative sample are input into the target risk control model to obtain the predicted value output by the target risk control model; Based on the predicted values, determine the area under the curve; The similarity score screening threshold corresponding to the largest area under the curve among the various areas under the curves is determined as the target quality screening threshold.

[0013] According to a risk control training sample optimization method provided by the present invention, the step of optimizing each candidate positive sample based on the target quality screening threshold includes: Candidate positive samples whose similarity scores are greater than or equal to the target quality screening threshold and less than the target preset threshold are assigned a first weight; Candidate positive samples whose similarity scores are less than the target quality screening threshold are assigned a second weight or are discarded.

[0014] The present invention also provides a risk control training sample optimization device, comprising: The rule extraction module is used to extract the feature paths corresponding to at least one target leaf node in the target decision tree model, and determine the feature paths as a feature combination rule; wherein, the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set includes multiple training samples of at least one dimension and label data of each training sample; the dimension is user behavior features and / or user transaction features; The sample extraction module is used to extract target training samples that satisfy each of the feature combination rules from the historical risk control data training set, and to determine each of the target training samples as candidate positive samples. The first determining module is used to determine, for each combination rule group, a similarity score for each candidate positive sample among all candidate positive samples corresponding to the combination rule group, based on the target Gaussian mixture model corresponding to the combination rule group; the target Gaussian mixture model is trained based on the first positive sample corresponding to the common features of the combination rule group and the candidate positive sample; the combination rule group includes at least two feature combination rules. An optimization module is used to optimize each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combined rule group.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the risk control training sample optimization method as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the risk control training sample optimization method as described above.

[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the risk control training sample optimization method as described above.

[0018] The present invention provides a risk control training sample optimization method, apparatus, electronic device, and storage medium. The method involves extracting feature paths corresponding to at least one target leaf node in a target decision tree model and defining these feature paths as feature combination rules. The target decision tree model is trained on a historical risk control data training set, which includes multiple training samples of at least one dimension and label data for each training sample. The dimension is user behavior features and / or user transaction features. Target training samples satisfying each feature combination rule are extracted from the historical risk control data training set, and each target training sample is defined as a candidate positive sample. For each combination rule group, a similarity score is determined for each candidate positive sample among all candidate positive samples corresponding to the combination rule group, based on the target Gaussian mixture model corresponding to the combination rule group. The target Gaussian mixture model is trained based on the first positive sample corresponding to the common feature of the combination rule group and the candidate positive samples. The combination rule group includes at least two feature combination rules. Based on the similarity score corresponding to each candidate positive sample in the combination rule group, each candidate positive sample is optimized. By combining the target decision tree model and the target Gaussian mixture model, and by scoring the similarity of each candidate positive sample in each combination rule group, the model optimizes each candidate positive sample, improves sample quality and business rationality, and thus enhances the performance of the model's risk control model. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts illustrating the risk control training sample optimization method provided by the present invention.

[0021] Figure 2 This is the second flowchart of the risk control training sample optimization method provided by the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of the risk control training sample optimization device provided by the present invention.

[0023] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] The following is combined with Figures 1-2 The present invention describes the risk control training sample optimization method.

[0026] Figure 1 This is one of the flowcharts illustrating the risk control training sample optimization method provided by this invention, such as... Figure 1 As shown, the method includes the following steps 101-104.

[0027] Step 101: Extract the feature path corresponding to at least one target leaf node in the target decision tree model, and determine the feature path as a feature combination rule; wherein, the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set includes multiple training samples of at least one dimension and label data of each training sample; the dimension is user behavior features and / or user transaction features.

[0028] It should be noted that the risk control training sample optimization method provided by the present invention can be applied to the scenario of training sample enhancement of risk control models in the field of risk control, for example, the risk control model is a black market identification model; the execution subject of the method can be a risk control training sample optimization device, such as an electronic device, or a control module in the risk control training sample optimization device for executing the risk control training sample optimization method.

[0029] Specifically, the target decision tree model is trained on a historical risk control data training set. This training set includes multiple training samples in at least one dimension and label data for each sample (e.g., whether it is a black market sample). The dimension represents user behavior characteristics and / or user transaction characteristics; for example, user behavior characteristics include login frequency, and user transaction characteristics include transaction amount, transaction device, or geographical location. The label data represents either abnormal samples (e.g., fraudulent samples, i.e., black market samples) or normal samples (e.g., non-fraudulent samples, i.e., non-black market samples).

[0030] A decision tree model with a clear structure and strong interpretability is generated by constructing split nodes using preset classification criteria (such as the Gini coefficient or information gain). The root node represents the initial split point of the entire sample space; intermediate nodes represent judgment conditions for further splitting samples according to a certain feature and its threshold; and leaf nodes represent the final sample subsets, containing the sample distribution under specific feature combinations. For example, the depth of the decision tree model is set to 5-8 layers to ensure rule interpretability, and the minimum number of samples per node is set to 500 to improve statistical significance.

[0031] After training the target decision tree model using historical risk control data training set, for each leaf node, the feature path from the root node to that leaf node is traced back, and all features and their value ranges on the feature path are extracted to form a feature combination rule. For example, "age > 50 and registration location = a certain region and login count > 5" is a feature combination rule.

[0032] Step 102: Extract target training samples that satisfy each of the feature combination rules from the historical risk control data training set, and determine each of the target training samples as candidate positive samples.

[0033] Specifically, based on the extracted feature combination rules, target training samples that satisfy each feature combination rule are extracted from the historical risk control data training set. That is, for each feature combination rule and its threshold conditions, the features and threshold conditions are compared with all training samples in the historical risk control data training set. If all feature values ​​of a training sample meet the judgment path of the feature combination rule, that is, although these training samples are currently marked as negative samples (label data is abnormal samples), their feature combination meets the high-risk behavior pattern. Therefore, the training sample is included in the candidate set corresponding to the feature combination rule and marked as a candidate positive sample.

[0034] Step 103: For each combination rule group, based on the target Gaussian mixture model corresponding to the combination rule group, determine the similarity score corresponding to each candidate positive sample among all candidate positive samples corresponding to the combination rule group; the target Gaussian mixture model is trained based on the first positive sample corresponding to the common feature of the combination rule group and the candidate positive sample; the combination rule group includes at least two feature combination rules.

[0035] Specifically, the first positive sample is the true positive sample, that is, the labeled data in the historical training dataset are normal training samples. When constructing the Gaussian Mixture Model (GMM) for each group, different numbers of components are set (e.g., from 1 to 10), where each component represents the number of clusters in the Gaussian Mixture Model. The corresponding Gaussian Mixture Models are trained separately, and the Bayesian Information Criterion (BIC) value is calculated. Finally, the Gaussian Mixture Model with the smallest BIC value is selected as the GMM model for that group, so as to achieve effective modeling of the candidate sample distribution.

[0036] The Gaussian mixture model corresponding to the combined rule group is trained based on the first positive sample (i.e. the true positive sample) and the candidate positive sample corresponding to the common features of the combined rule group. The parameter distribution is estimated by the expectation-maximization (EM) algorithm to ensure that the Gaussian mixture model can better capture the potential structure of the "true positive class distribution" under the combined rule group.

[0037] After training to obtain the target Gaussian mixture model, based on the target Gaussian mixture model corresponding to the combination rule group, the similarity score corresponding to each candidate positive sample among all candidate positive samples corresponding to the combination rule group can be determined.

[0038] By training a Gaussian mixture model based on real positive samples, the similarity between the distribution of candidate positive samples and the distribution of real positive samples can be accurately quantified. Compared with traditional oversampling techniques, this method can better capture the true distribution characteristics of the data, and is especially suitable for multimodal data.

[0039] Step 104: Optimize each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combined rule group.

[0040] Specifically, based on the similarity score corresponding to each candidate positive sample in the combination rule group, each candidate positive sample can be further optimized.

[0041] The risk control training sample optimization method provided by this invention extracts feature paths corresponding to at least one target leaf node in a target decision tree model and determines the feature paths as feature combination rules. The target decision tree model is trained on a historical risk control data training set, which includes multiple training samples of at least one dimension and label data for each training sample. The dimension is user behavior features and / or user transaction features. Target training samples satisfying each feature combination rule are extracted from the historical risk control data training set, and each target training sample is determined as a candidate positive sample. For each combination rule group, a similarity score is determined for each candidate positive sample among all candidate positive samples corresponding to the combination rule group, based on the target Gaussian mixture model corresponding to the combination rule group. The target Gaussian mixture model is trained based on the first positive sample corresponding to the common feature of the combination rule group and the candidate positive samples. The combination rule group includes at least two feature combination rules. Based on the similarity score corresponding to each candidate positive sample in the combination rule group, each candidate positive sample is optimized. By combining the target decision tree model and the target Gaussian mixture model, and by scoring the similarity of each candidate positive sample in each combination rule group, the model optimizes each candidate positive sample, improves sample quality and business rationality, and thus enhances the performance of the model's risk control model.

[0042] Optionally, the specific implementation of step 101 above includes: Traverse each leaf node of the target decision tree model to determine the sample set corresponding to each leaf node; each leaf node includes at least one positive sample; the positive sample is determined based on the label data of each training sample; the sample set includes the at least one positive sample; based on the number of each positive sample in the sample set, determine the at least one target leaf node from each leaf node; based on each target leaf node, determine the feature path corresponding to each target leaf node respectively.

[0043] Specifically, the trained target decision tree model is traversed to obtain the sample set corresponding to each leaf node. Each leaf node includes at least one positive sample. Positive samples are determined based on the label data of each training sample, and the sample set includes at least one positive sample.

[0044] The number of positive samples in the sample set corresponding to each leaf node and the total number of samples are counted to determine the proportion of positive samples in each leaf node. If the proportion of positive samples in a leaf node exceeds a first preset threshold (e.g., 30%) and the total number of samples in the leaf node exceeds a second preset threshold (e.g., 200), then the leaf node is determined as the target leaf node. The first and second preset thresholds can be adjusted according to the actual situation and are not limited in any way.

[0045] Based on the target leaf node, the feature path from the root node to the target leaf node in the target decision tree model is traced back, and the feature path corresponding to the leaf node is extracted as a high-confidence feature combination rule. For example, "age < 30 and registration location = a certain region and login count > 5" is a typical feature combination rule.

[0046] By selecting feature combinations that have a higher proportion of positive samples than a first preset threshold and cover a large number of training samples, these rules serve as interpretability rules that reflect high-risk behaviors. These rules reflect high-risk patterns in actual business and can effectively screen potential candidate positive samples.

[0047] Optionally, the method further includes: Based on any two feature combination rules among all feature combination rules, determine the number of common features between the feature sets involved in the two feature combination rules; based on the number of common features, determine the feature overlap between the two feature combination rules; based on the feature overlap, construct a similarity matrix; based on the similarity matrix, group all feature combination rules to obtain at least one combination rule group; for each combination rule group, extract the common features of all combination rules in the combination rule group.

[0048] Specifically, based on any two feature combination rules among all feature combination rules, the feature sets involved are extracted respectively, the number of common features between the feature sets involved in any two feature combination rules is counted, and the number of common features is divided by the total number of the union of the feature sets involved in any two feature combination rules to obtain the feature overlap between any two feature combination rules.

[0049] For example, if feature combination rule A uses {age, region, number of logins} and feature combination rule B uses {region, number of logins, device type}, then the common feature of feature combination rules A and B is {region, number of logins}, and the feature union is {age, region, number of logins, device type}. Therefore, the overlap is 2 / 4 = 0.5.

[0050] Based on the overlap of each feature, a similarity matrix is ​​constructed. Using this similarity matrix, the Average Linkage method from hierarchical clustering is employed to aggregate all feature combination rules from the bottom up. A similarity threshold is set during the clustering process, for example, 0.6. Two feature combination rules with a feature overlap greater than the similarity threshold are grouped into a single group, resulting in at least one group. Each group includes at least two highly similar feature combination rules to facilitate focusing on the local feature space during subsequent modeling. For example, each group may include 3-5 highly similar feature combination rules to ensure sufficient diversity during subsequent Gaussian mixture modeling without excessive model complexity due to an excessive number of feature combination rules.

[0051] For each set of combined rules, extract the common features of all combined rules in the set to form a feature subset.

[0052] Optionally, the specific implementation process of step 103 above includes: Each candidate positive sample in the group of combined rules is input into the target Gaussian mixture model to obtain the log-likelihood value of each candidate positive sample output by the target Gaussian mixture model; the log-likelihood value of each candidate positive sample in the group of combined rules is normalized to obtain the similarity score of each candidate positive sample in the group of combined rules.

[0053] Specifically, by inputting each candidate positive sample from all candidate positive samples corresponding to the combination rule group into the target Gaussian mixture model, the log-likelihood value corresponding to each candidate positive sample output by the target Gaussian mixture model can be obtained. This log-likelihood value reflects the probability density of the candidate positive sample under the "positive class distribution" of the group. The larger the value, the closer it is to the distribution of the real positive sample.

[0054] For each combination rule group, the log-likelihood values ​​of all candidate positive samples are normalized (e.g., Min-Max mapping) and mapped to the [0,1] interval, which serves as the similarity score for each candidate sample, used for subsequent screening of high-quality samples or to assist in model training.

[0055] Optionally, optimizing each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combination rule group includes: Based on the similarity score corresponding to each candidate positive sample in the combined rule group, a target quality screening threshold is determined; based on the target quality screening threshold, each candidate positive sample is optimized.

[0056] Specifically, based on the similarity score corresponding to each candidate positive sample in the combined rule group, the target quality screening threshold can be determined; for example, the target quality screening threshold is 0.7. Then, based on the target quality screening threshold, each candidate positive sample is screened and weighted.

[0057] Optionally, determining the target quality screening threshold based on the similarity score corresponding to each candidate positive sample in the combined rule set includes: Based on the second positive sample and the first negative sample in the validation set, and the similarity score corresponding to each candidate positive sample, multiple similarity score screening thresholds are tested within the target range; the target quality screening threshold is determined from each of the similarity score screening thresholds.

[0058] Specifically, the validation set includes a second positive sample and a first negative sample, wherein both the second positive sample and the first negative sample have real labeled data. The labeled data of the second positive sample is normal sample (e.g., non-fraudulent sample), and the labeled data of the first negative sample is abnormal sample (e.g., fraudulent sample).

[0059] Based on the second positive sample and the first negative sample in the validation set, and the similarity score (range [0,1]) corresponding to each candidate positive sample, multiple similarity score screening thresholds are tested within the target range (e.g., 0.5-0.8). Specifically, the second positive sample and the first negative sample are input into the target Gaussian mixture model to obtain the log-likelihood values ​​for the second positive sample and the first negative sample, respectively. The similarity scores for the second positive sample and the first negative sample are then determined based on these values. The similarity scores for each candidate positive sample, the second positive sample, and the first negative sample are used as multiple similarity score screening thresholds. Alternatively, the highest similarity score among these three scores can be normalized and used as a single similarity score threshold. Finally, the target quality screening threshold is determined from these similarity score screening thresholds.

[0060] Optionally, determining the target quality screening threshold from each of the similarity score screening thresholds includes: Based on each of the similarity score screening thresholds, candidate positive samples corresponding to similarity scores greater than the similarity score screening thresholds are determined as expanded training samples; the expanded training samples are used to train an initial risk control model to obtain a target risk control model; the second positive sample and the first negative sample are input into the target risk control model to obtain the predicted value output by the target risk control model; based on the predicted value, the area under the curve is determined; the similarity score screening threshold corresponding to the largest area under the curve among the various areas under the curve is determined as the target quality screening threshold.

[0061] Specifically, for each similarity score screening threshold, candidate positive samples corresponding to similarity scores greater than the similarity score screening threshold can be determined as expanded training samples; the initial risk control model can be trained using the expanded training samples to obtain the target risk control model; the second positive sample and the first negative sample in the validation set are input into the target risk control model to obtain the predicted value output by the target risk control model; based on the predicted value, the area under the curve (AUC) is determined; the similarity score screening threshold corresponding to the largest area under the curve among all AUCs is determined as the target quality screening threshold, that is, the similarity score screening threshold that makes the target risk control model have the largest AUC on the validation set is finally selected as the final target quality screening threshold.

[0062] Optionally, optimizing each of the candidate positive samples based on the target quality screening threshold includes: Candidate positive samples whose similarity scores are greater than or equal to the target quality screening threshold and less than the target preset threshold are assigned a first weight; candidate positive samples whose similarity scores are less than the target quality screening threshold are assigned a second weight or are discarded.

[0063] Specifically, this application provides a tiered weight allocation mechanism based on similarity scores. Candidate positive samples with similarity scores greater than or equal to a target quality screening threshold but less than a target preset threshold are assigned a first weight. For example, if the target quality screening threshold is 0.7 and the target preset threshold is 0.8, candidate positive samples with similarity scores greater than or equal to 0.7 but less than 0.8 are assigned a first weight of 0.7. Candidate positive samples with similarity scores greater than 0.8 are assigned a target weight of 1. The label data of the candidate positive samples with the assigned first and target weights is then modified to normal, i.e., a label value of 1, as the final sample augmentation result.

[0064] Candidate positive samples with similarity scores less than the target quality screening threshold are assigned a second weight or discarded. For example, candidate positive samples with similarity scores less than 0.7 are assigned a second weight of 0.3 or discarded according to the actual situation, so as to achieve dynamic control and rational utilization of candidate sample quality.

[0065] For example, with 100 candidate positive samples, each corresponding to a similarity score between 0 and 1, candidate positive samples with a similarity score greater than or equal to 0.8 have a weight of 1, candidate positive samples with a similarity score greater than or equal to 0.7 and less than 0.8 have a weight of 0.7, and candidate positive samples with a similarity score less than 0.7 may not be included as supplementary positive samples or may be assigned a low weight of 0.3, thus achieving dynamic control and rational utilization of candidate sample quality. The risk control model tends to favor samples with higher weights in predicting correctly; therefore, true positive samples generally have the highest weight, samples very similar to true positive samples have lower weights, and generally, similar samples have the lowest weight.

[0066] By employing a tiered weighting mechanism based on similarity scores, different weights are assigned to candidate positive samples of varying quality levels, thereby improving the quality and stability of model training. Compared to existing fixed-weight or simple linear weighting methods, this approach is better suited to complex data environments and effectively reduces the negative impact of low-quality samples on the model.

[0067] Figure 2 This is the second flowchart of the risk control training sample optimization method provided by the present invention, as shown below. Figure 2As shown, it includes feature combination rule extraction and candidate positive sample generation of decision tree model, feature combination rule grouping, grouped Gaussian mixture model modeling and scoring, and dynamic weight allocation strategy; wherein, feature combination rule extraction and candidate positive sample generation of decision tree model: (1) Decision tree model training. Split nodes are constructed by pre-set classification criteria (such as Gini coefficient or information gain) to generate a decision tree model with clear structure and strong interpretability. The decision tree model is trained based on historical risk control data training set to obtain target decision tree model, wherein the historical risk control data training set includes multiple training samples of at least one dimension and label data of each training sample; the dimension is user behavior features and / or user transaction features, such as user behavior features being login frequency, and user transaction features being transaction amount, transaction device or geographical location. The label data is abnormal samples (such as fraud samples) or normal samples (such as non-fraud samples). After the target decision tree model is trained using historical risk control data training set, for each leaf node, the feature path from the root node to the leaf node is traced back, and all features and their value ranges on the feature path are extracted to form a feature combination rule. For example, “age > 50 and registered location = a certain region and login times > 5” is a feature combination rule. (2) Determine the positive sample ratio and total number of samples of the leaf node. Traverse the trained target decision tree model to obtain the sample set corresponding to each leaf node. The leaf node includes at least one positive sample. The positive sample is determined based on the label data of each training sample. The sample set includes at least one positive sample. Count the number of positive samples in the sample set corresponding to each leaf node and the total number of samples to determine the positive sample ratio of each leaf node. If the positive sample ratio in the leaf node exceeds the first preset threshold (e.g., 30%) and the total number of samples in the leaf node exceeds the second preset threshold (e.g., 200), then the leaf node is determined as the target leaf node. Based on the target leaf node, trace back the feature path from the root node to the target leaf node in the target decision tree model and extract the feature path corresponding to the leaf node as a high-confidence feature combination rule. (3) Mark candidate positive samples. Based on the extracted feature combination rules, target training samples that satisfy each feature combination rule are extracted from the historical risk control data training set. That is, for each feature combination rule and its threshold conditions, the features and threshold conditions are compared with all training samples in the historical risk control data training set. If all feature values ​​of a training sample meet the judgment path of the feature combination rule, that is, although these training samples are currently marked as negative samples (label data is abnormal samples), their feature combination meets the high-risk behavior pattern. Therefore, the training sample is included in the candidate set corresponding to the feature combination rule and marked as a candidate positive sample.

[0068] Feature combination rule grouping: (1) Feature overlap calculation. Based on any two feature combination rules in all feature combination rules, extract the feature sets involved respectively, count the number of common features between the feature sets involved in any two feature combination rules, and divide the number of common features by the total number of the union of the feature sets involved in any two feature combination rules to obtain the feature overlap between any two feature combination rules. (2) Hierarchical clustering. Based on each feature overlap, construct a similarity matrix; based on the similarity matrix, use the average linking method in hierarchical clustering to aggregate all feature combination rules from bottom to top, and set a similarity threshold during the clustering process, for example, a similarity threshold of 0.6. (3) Combination rule group. Divide two feature combination rules with feature overlap greater than the similarity threshold into a combination rule group to obtain at least one combination rule group. Each combination rule group includes at least two feature combination rules, and at least two feature combination rules are highly similar feature combination rules. For each combination rule group, extract the common features of all combination rules in the combination rule group to form a feature subset.

[0069] Grouped Gaussian Mixture Model Modeling and Scoring: (1) Grouped Gaussian Mixture Model Training. When constructing the Gaussian Mixture Model (GMM) for each group, different numbers of components are set (e.g., from 1 to 10), and the corresponding Gaussian Mixture Models are trained respectively. The Bayesian Information Criterion (BIC) value is calculated, and the Gaussian Mixture Model with the smallest BIC value is finally selected as the GMM model for that group to achieve effective modeling of the candidate sample distribution. The Gaussian Mixture Model corresponding to the combined rule group is trained based on the first positive sample and the candidate positive sample corresponding to the common features of the combined rule group. The parameter distribution is estimated by the expectation-maximization (EM) algorithm to ensure that the Gaussian Mixture Model better captures the potential structure of the "true positive class distribution" under the combined rule group. (2) Calculation of Log-Likelihood Value of Candidate Positive Samples. Input each candidate positive sample from all candidate positive samples corresponding to the combined rule group into the target Gaussian mixture model to obtain the log-likelihood value of each candidate positive sample output by the target Gaussian mixture model; the log-likelihood value reflects the probability density of the candidate positive sample under the "positive class distribution" of the group, and the larger the value, the closer it is to the distribution of the real positive sample. (3) Log-likelihood value normalization. Normalize the log-likelihood values ​​of all candidate positive samples corresponding to each combined rule group (such as Min-Max mapping), map the log-likelihood values ​​to the [0,1] interval, and use them as the similarity score of each candidate sample for subsequent screening of high-quality samples or auxiliary model training.

[0070] Dynamic weight allocation strategy: (1) Determine the target quality screening threshold. Based on the second positive sample and the first negative sample in the validation set, and the similarity score (range [0,1]) corresponding to each candidate positive sample, test multiple similarity score screening thresholds within the target range (e.g., 0.5-0.8). For each similarity score screening threshold, the candidate positive samples corresponding to similarity scores greater than the similarity score screening threshold can be determined as the expanded training samples; the expanded training samples are used to train the initial risk control model to obtain the target risk control model; the second positive sample and the first negative sample in the validation set are input into the target risk control model to obtain the predicted value output by the target risk control model; based on the predicted value, the area under the curve (AUC) is determined; the similarity score screening threshold corresponding to the largest area under the curve among each area under the curve is determined as the target quality screening threshold, that is, the similarity score screening threshold that makes the target risk control model have the largest AUC on the validation set is finally selected as the final target quality screening threshold. (2) Sample screening and weighting. A tiered weighting mechanism based on similarity scores is used. Candidate positive samples with similarity scores greater than or equal to the target quality screening threshold but less than a preset target threshold are assigned a first weight. For example, candidate positive samples with similarity scores greater than or equal to 0.7 but less than 0.8 are assigned a first weight of 0.7, and candidate positive samples with similarity scores greater than or equal to 0.8 are assigned a target weight of 1. The label data of the candidate positive samples with the assigned first and target weights is then modified to normal, i.e., a label value of 1, as the final sample augmentation result. Candidate positive samples with similarity scores less than the target quality screening threshold are assigned a second weight or discarded. For example, candidate positive samples with similarity scores less than 0.7 are assigned a second weight of 0.3 or discarded based on actual circumstances, thus achieving dynamic control and rational utilization of candidate sample quality.

[0071] The risk control training sample optimization device provided by the present invention is described below. The risk control training sample optimization device described below and the risk control training sample optimization method described above can be referred to in correspondence.

[0072] Figure 3 This is a schematic diagram of the structure of the risk control training sample optimization device provided by the present invention, as shown below. Figure 3 As shown, the risk control training sample optimization device 300 includes: a rule extraction module 301, a sample extraction module 302, a first determination module 303, and an optimization module 304; wherein, The rule extraction module 301 is used to extract the feature paths corresponding to at least one target leaf node in the target decision tree model, and determine the feature paths as a feature combination rule; wherein, the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set includes multiple training samples of at least one dimension and label data of each training sample; the dimension is user behavior features and / or user transaction features; The sample extraction module 302 is used to extract target training samples that satisfy each of the feature combination rules from the historical risk control data training set, and to determine each of the target training samples as candidate positive samples; The first determining module 303 is used to determine, for each combination rule group, a similarity score corresponding to each candidate positive sample among all candidate positive samples corresponding to the combination rule group, based on the target Gaussian mixture model corresponding to the combination rule group; the target Gaussian mixture model is trained based on the first positive sample corresponding to the common features of the combination rule group and the candidate positive sample; the combination rule group includes at least two feature combination rules. The optimization module 304 is used to optimize each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combination rule group.

[0073] The risk control training sample optimization device provided by this invention extracts feature paths corresponding to at least one target leaf node in a target decision tree model and determines the feature paths as feature combination rules. The target decision tree model is trained on a historical risk control data training set, which includes multiple training samples of at least one dimension and label data for each training sample. The dimension is user behavior features and / or user transaction features. Target training samples satisfying each feature combination rule are extracted from the historical risk control data training set, and each target training sample is determined as a candidate positive sample. For each combination rule group, a similarity score is determined for each candidate positive sample among all candidate positive samples corresponding to the combination rule group, based on the target Gaussian mixture model corresponding to the combination rule group. The target Gaussian mixture model is trained based on the first positive sample corresponding to the common feature of the combination rule group and the candidate positive samples. The combination rule group includes at least two feature combination rules. Each candidate positive sample is optimized based on its similarity score. By combining the target decision tree model and the target Gaussian mixture model, and by scoring the similarity of each candidate positive sample in each combination rule group, the model optimizes each candidate positive sample, improves sample quality and business rationality, and thus enhances the performance of the model's risk control model.

[0074] Optionally, the rule extraction module 301 is specifically used for: Traverse each leaf node of the target decision tree model to determine the sample set corresponding to each leaf node; each leaf node includes at least one positive sample; the positive sample is determined based on the label data of each training sample; the sample set includes the at least one positive sample. Based on the number of each of the positive samples in the sample set, at least one target leaf node is determined from each of the leaf nodes; Based on each of the target leaf nodes, the feature path corresponding to each target leaf node is determined.

[0075] Optionally, the risk control training sample optimization device 300 further includes: The second determining module is used to determine the number of common features between the feature sets involved in any two feature combination rules based on any two feature combination rules among all feature combination rules; The third determining module is used to determine the feature overlap between any two feature combination rules based on the number of common features; A construction module is used to construct a similarity matrix based on the overlap of each of the aforementioned features; The grouping module is used to group all feature combination rules based on the similarity matrix to obtain at least one combination rule group; The common feature extraction module is used to extract the common features of all combined rules in each combined rule group.

[0076] Optionally, the first determining module 303 is specifically used for: Each candidate positive sample and the first sample in all candidate positive samples corresponding to the combined rule group are input into the target Gaussian mixture model to obtain the log likelihood value corresponding to each candidate positive sample output by the target Gaussian mixture model. The log-likelihood values ​​of all candidate positive samples corresponding to each of the combined rule groups are normalized to obtain the similarity score of each candidate positive sample among all candidate positive samples corresponding to the combined rule groups.

[0077] Optionally, the optimization module 304 is specifically used for: Based on the similarity score corresponding to each candidate positive sample in the combined rule group, the target quality screening threshold is determined; Based on the target quality screening threshold, each of the candidate positive samples is optimized.

[0078] Optionally, the optimization module 304 is further configured to: Based on the second positive sample and the first negative sample in the validation set, and the similarity score corresponding to each candidate positive sample, multiple similarity score screening thresholds are tested within the target range. The target quality screening threshold is determined from each of the aforementioned similarity score screening thresholds.

[0079] Optionally, the optimization module 304 is further configured to: Based on each of the similarity score screening thresholds, the candidate positive samples corresponding to similarity scores greater than the similarity score screening thresholds are determined as the expanded training samples. The expanded training samples are used to train the initial risk control model to obtain the target risk control model. The second positive sample and the first negative sample are input into the target risk control model to obtain the predicted value output by the target risk control model; Based on the predicted values, determine the area under the curve; The similarity score screening threshold corresponding to the largest area under the curve among the various areas under the curves is determined as the target quality screening threshold.

[0080] Optionally, the optimization module 304 is specifically used for: Candidate positive samples whose similarity scores are greater than or equal to the target quality screening threshold and less than the target preset threshold are assigned a first weight; Candidate positive samples whose similarity scores are less than the target quality screening threshold are assigned a second weight or are discarded.

[0081] Figure 4 This is a schematic diagram of the physical structure of an electronic device provided by the present invention, such as... Figure 4As shown, the electronic device 400 may include: a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communications bus 440. The processor 410 can call logical instructions in the memory 430 to execute a risk control training sample optimization method. This method includes: extracting feature paths corresponding to at least one target leaf node in the target decision tree model, and determining the feature paths as feature combination rules; wherein the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set including multiple training samples of at least one dimension and label data for each training sample; the dimension being user behavior features and / or user transaction features; extracting target training samples satisfying each of the feature combination rules from the historical risk control data training set, and determining each target training sample as a candidate positive sample; for each combination rule group, determining a similarity score for each candidate positive sample among all candidate positive samples corresponding to the combination rule group based on the target Gaussian mixture model corresponding to the combination rule group; the target Gaussian mixture model is trained based on a first positive sample corresponding to a common feature of the combination rule group and the candidate positive samples; the combination rule group includes at least two feature combination rules; and optimizing each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combination rule group.

[0082] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the risk control training sample optimization method provided by the above methods. The method includes: extracting feature paths corresponding to at least one target leaf node in the target decision tree model, and determining the feature path as a feature combination rule; wherein the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set including multiple training samples of at least one dimension and label data of each training sample; the dimension is user behavior features and / or User transaction characteristics; extract target training samples that satisfy each of the feature combination rules from the historical risk control data training set, and determine each of the target training samples as candidate positive samples; for each combination rule group, determine the similarity score corresponding to each candidate positive sample in all candidate positive samples corresponding to the combination rule group based on the target Gaussian mixture model corresponding to the combination rule group; the target Gaussian mixture model is trained based on the first positive sample corresponding to the common feature of the combination rule group and the candidate positive samples; the combination rule group includes at least two of the feature combination rules; optimize each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combination rule group.

[0084] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the risk control training sample optimization method provided by the above methods. The method includes: extracting feature paths corresponding to at least one target leaf node in a target decision tree model, and determining the feature path as a feature combination rule; wherein the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set including multiple training samples of at least one dimension and label data for each training sample; the dimension being user behavior features and / or user transaction features; and ... Target training samples satisfying the aforementioned feature combination rules are extracted from the risk control data training set, and each of the aforementioned target training samples is determined as a candidate positive sample; for each combination rule group, based on the target Gaussian mixture model corresponding to the combination rule group, the similarity score corresponding to each of the candidate positive samples in all candidate positive samples corresponding to the combination rule group is determined; the target Gaussian mixture model is trained based on the first positive sample corresponding to the common feature of the combination rule group and the candidate positive sample; the combination rule group includes at least two of the aforementioned feature combination rules; based on the similarity score corresponding to each of the candidate positive samples in the combination rule group, each of the candidate positive samples is optimized.

[0085] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0086] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing risk control training samples, characterized in that, include: Feature paths corresponding to at least one target leaf node in the target decision tree model are extracted, and the feature paths are determined as a feature combination rule; wherein, the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set includes multiple training samples of at least one dimension and label data of each training sample; the dimension is user behavior features and / or user transaction features; user behavior features are login frequency, and user transaction features are transaction amount, transaction device, or geographical location; Target training samples that satisfy each of the feature combination rules are extracted from the historical risk control data training set, and each of the target training samples is determined as a candidate positive sample; by selecting feature combination rules that have a positive sample ratio higher than a first preset threshold and cover a large number of training samples, they are used as interpretable rules that reflect high-risk behavior. These rules reflect high-risk patterns in actual business and effectively screen out potential candidate positive samples. For each set of combined rules, based on the target Gaussian mixture model corresponding to the combined rule set, a similarity score is determined for each candidate positive sample among all candidate positive samples corresponding to the combined rule set; the target Gaussian mixture model is trained based on the first positive sample corresponding to the common feature of the combined rule set and the candidate positive sample; the combined rule set includes at least two feature combination rules. Based on the similarity score corresponding to each candidate positive sample in the combined rule group, each candidate positive sample is optimized; Based on any two feature combination rules among all feature combination rules, determine the number of common features between the feature sets involved in the two feature combination rules. Based on the number of common features, determine the feature overlap between any two feature combination rules; Based on the overlap of each feature, a similarity matrix is ​​constructed; Based on the similarity matrix, all feature combination rules are grouped to obtain at least one combination rule group; For each group of combined rules, extract the common features of all combined rules in the group.

2. The risk control training sample optimization method according to claim 1, characterized in that, The step of extracting the feature path corresponding to at least one target leaf node in the target decision tree model includes: Traverse each leaf node of the target decision tree model to determine the sample set corresponding to each leaf node; each leaf node includes at least one positive sample; the positive sample is determined based on the label data of each training sample; the sample set includes the at least one positive sample. Based on the number of each of the positive samples in the sample set, at least one target leaf node is determined from each of the leaf nodes; Based on each of the target leaf nodes, the feature path corresponding to each target leaf node is determined.

3. The risk control training sample optimization method according to claim 1, characterized in that, The step of determining the similarity score for each candidate positive sample among all candidate positive samples corresponding to the combined rule group, based on the target Gaussian mixture model corresponding to the combined rule group, includes: Each candidate positive sample in the group of combined rules is input into the target Gaussian mixture model to obtain the log-likelihood value of each candidate positive sample output by the target Gaussian mixture model. The log-likelihood values ​​of all candidate positive samples corresponding to each of the combined rule groups are normalized to obtain the similarity score of each candidate positive sample among all candidate positive samples corresponding to the combined rule groups.

4. The risk control training sample optimization method according to claim 1, characterized in that, The optimization of each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combined rule group includes: Based on the similarity score corresponding to each candidate positive sample in the combined rule group, the target quality screening threshold is determined; Based on the target quality screening threshold, each of the candidate positive samples is optimized.

5. The risk control training sample optimization method according to claim 4, characterized in that, The step of determining the target quality screening threshold based on the similarity score corresponding to each candidate positive sample in the combined rule group includes: Based on the second positive sample and the first negative sample in the validation set, and the similarity score corresponding to each candidate positive sample, multiple similarity score screening thresholds are tested within the target range. The target quality screening threshold is determined from each of the aforementioned similarity score screening thresholds.

6. The risk control training sample optimization method according to claim 5, characterized in that, Determining the target quality screening threshold from each of the similarity score screening thresholds includes: Based on each of the similarity score screening thresholds, the candidate positive samples corresponding to similarity scores greater than the similarity score screening thresholds are determined as the expanded training samples. The expanded training samples are used to train the initial risk control model to obtain the target risk control model. The second positive sample and the first negative sample are input into the target risk control model to obtain the predicted value output by the target risk control model; Based on the predicted values, determine the area under the curve; The similarity score screening threshold corresponding to the largest area under the curve among the various areas under the curves is determined as the target quality screening threshold.

7. The risk control training sample optimization method according to claim 4, characterized in that, The optimization of each candidate positive sample based on the target quality screening threshold includes: Candidate positive samples whose similarity scores are greater than or equal to the target quality screening threshold and less than the target preset threshold are assigned a first weight; Candidate positive samples whose similarity scores are less than the target quality screening threshold are assigned a second weight or are discarded.

8. A risk control training sample optimization device, characterized in that, include: The rule extraction module is used to extract the feature paths corresponding to at least one target leaf node in the target decision tree model, and determine the feature paths as a feature combination rule; wherein, the target decision tree model is trained based on a historical risk control data training set, the historical risk control data training set includes multiple training samples of at least one dimension and label data of each training sample; the dimension is user behavior features and / or user transaction features; the user behavior features are login frequency, and the user transaction features are transaction amount, transaction device, or geographical location; The sample extraction module is used to extract target training samples that satisfy each of the feature combination rules from the historical risk control data training set, and to determine each of the target training samples as candidate positive samples; by selecting feature combination rules that have a positive sample ratio higher than a first preset threshold and cover a large number of training samples, these rules are used as interpretable rules that reflect high-risk behavior. These rules reflect high-risk patterns in actual business and effectively screen out potential candidate positive samples. The first determining module is used to determine, for each combination rule group, a similarity score for each candidate positive sample among all candidate positive samples corresponding to the combination rule group, based on the target Gaussian mixture model corresponding to the combination rule group; the target Gaussian mixture model is trained based on the first positive sample corresponding to the common features of the combination rule group and the candidate positive sample; the combination rule group includes at least two feature combination rules. An optimization module is used to optimize each candidate positive sample based on the similarity score corresponding to each candidate positive sample in the combination rule group; The second determining module is used to determine the number of common features between the feature sets involved in any two feature combination rules based on any two feature combination rules among all feature combination rules; The third determining module is used to determine the feature overlap between any two feature combination rules based on the number of common features; A construction module is used to construct a similarity matrix based on the overlap of each of the aforementioned features; The grouping module is used to group all feature combination rules based on the similarity matrix to obtain at least one combination rule group; The common feature extraction module is used to extract the common features of all combined rules in each combined rule group.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the risk control training sample optimization method as described in any one of claims 1 to 7.

Images