Fusion training method and system for risk control model based on knowledge representation learning

Abstract

The present disclosure provides a fusion training method for a risk control model, including: receiving label data and refining expert knowledge; performing multi-order feature crossover on the label data and the expert knowledge to obtain data representation and rule representation; purifying the data representation based on the rule representation; and training and outputting the risk control model based on the purified data representation.

Description

Technical Field

[0001] The present disclosure mainly relates to knowledge representation learning, and in particular to risk control model training based on knowledge representation learning. Background Art

[0002] To avoid transaction event risks, the goal of risk control trust is to find risk-free pure white traffic for quick release. The accumulation of trusted data can help release low-risk transaction events and reduce the amount of analysis at the identification layer. Conventional risk trust models use pre-defined black and white samples for trust model training. Black samples come from the review of payment events complained by users; white samples come from events where users have successfully paid and do not involve complaints, audits, management and other risk control actions. Compared with the huge amount of white samples, black samples are very likely to be insufficient to reflect the full picture of risk events. Insufficient black samples usually lead to insufficient robustness of risk control trust models.

[0003] Therefore, this field needs an efficient risk control model training method that can improve the robustness of the model. Summary of the invention

[0004] To solve the above technical problems, the present disclosure provides a fusion training scheme based on knowledge representation learning for a risk control model, which is based on the accumulation of expert experience in the field of risk control, and improves the robustness of the risk control model by incorporating multi-order feature intersection and data purification, and makes the interpretability of the risk control model meet the requirements.

[0005] In one embodiment of the present disclosure, a fusion training method based on knowledge representation learning of a risk control model is provided, including: receiving labeled data and refining expert knowledge; performing multi-order feature crossover on the labeled data and expert knowledge to obtain data representation and rule representation respectively; purifying data representation based on rule representation; and training and outputting a risk control model based on the purified data representation.

[0006] In another embodiment of the present disclosure, the multi-order feature crossover includes a first-order feature crossover, a second-order feature crossover and a high-order feature crossover.

[0007] In yet another embodiment of the present disclosure, multi-order feature crossover of label data and expert knowledge is implemented through a data encoder and a rule encoder respectively.

[0008] In another embodiment of the present disclosure, purifying the data representation based on the rule representation further includes adopting a decision block composed of a plurality of expert blocks to purify the data representation.

[0009] In yet another embodiment of the present disclosure, purifying the data representation based on the rule representation includes introducing a rule-related loss function based on the rule representation.

[0010] In another embodiment of the present disclosure, purifying data representation based on rule representation includes constructing a fusion loss function of rule-related loss functions and task-related loss functions.

[0011] In yet another embodiment of the present disclosure, the label data is black and white label data.

[0012] In another embodiment of the present disclosure, the first-order feature crossover adopts a multi-layer perceptron MLP, the second-order feature crossover adopts a factorization machine FM, and the high-order feature crossover adopts a logarithmic neural network LNN.

[0013] In yet another embodiment of the present disclosure, using a decision block composed of a plurality of expert blocks to purify data representation is achieved through expert blocks with different weights.

[0014] In another embodiment of the present disclosure, training the risk control model based on the purified data representation includes optimizing the constructed fusion loss function.

[0015] In one embodiment of the present disclosure, a fusion training system based on knowledge representation learning of a risk control model is provided, including: an information acquisition module, which receives label data and refines expert knowledge; a feature crossover module, which performs multi-order feature crossover on label data and expert knowledge to obtain data representation and rule representation respectively; a purification module, which purifies data representation based on rule representation; and a training module, which trains and outputs a risk control model based on the purified data representation.

[0016] In one embodiment of the present disclosure, a computer-readable storage medium storing instructions is provided. When the instructions are executed, the machine executes the above method.

[0017] This Summary is provided to introduce some concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above invention content and the following detailed description of the present disclosure will be better understood when read in conjunction with the accompanying drawings. It should be noted that the accompanying drawings are only examples of the claimed invention. In the accompanying drawings, the same reference numerals represent the same or similar elements.

[0019] Figure 1 is a flow chart showing a fusion training method based on knowledge representation learning of a risk control model according to an embodiment of the present disclosure;

[0020] Figure 2 is a schematic diagram showing a model fusion training framework based on knowledge representation learning according to an embodiment of the present disclosure;

[0021] Figure 3 is a schematic diagram showing a multi-order feature crossover process of label data and expert knowledge for a risk control model according to an embodiment of the present disclosure;

[0022] Figure 4 is a schematic diagram showing a feature cross-implementation framework in a risk control scenario according to an embodiment of the present disclosure;

[0023] Figure 5 is a schematic diagram showing a data purification and model training process in fusion training based on knowledge representation learning of a risk control model according to an embodiment of the present disclosure;

[0024] Figure 6 is a block diagram showing a fusion training system based on knowledge representation learning of a risk control model according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0025] In order to make the above-mentioned objects, features and advantages of the present disclosure more obvious and understandable, the specific embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0026] In the following description, many specific details are set forth to facilitate a full understanding of the present disclosure. However, the present disclosure may also be implemented in other ways different from those described herein. Therefore, the present disclosure is not limited to the specific embodiments disclosed below.

[0027] In today's electronic payment environment, transaction events often contain risks. The goal of risk control trust is to find risk-free pure white traffic and quickly release it, which can reduce the disturbance to users on the one hand and save system computing resources on the other hand. The goal of global trust is to quickly release risk-free pure white traffic in all risk domains (for example, account theft).

[0028] Conventional trust models are modeled for predefined black and white samples. Black samples come from the review of payment events complained by users, and then the events confirmed to be involved are selected as black samples; white samples come from events in which users successfully pay and do not involve complaints, audits, management and other risk control actions. White samples are sometimes further processed, such as multiple successful transactions in a short period of time in the dimension of the user's active and passive relationship, but they are not separated from the comparison task with the black samples produced based on the qualitative nature of complaints.

[0029] Generally speaking, there is a problem of insufficient robustness when directly using black and white labeled data for trusted model training. This is because the risk concentrations in different risk domains are inconsistent. The risk control system can intercept most risk transactions in the risk domains that were built earlier, resulting in the magnitude of risk events ultimately exposed in the complaint samples being insufficient to support the training process of the trusted model. For example, a risk domain has less than one hundred risk event samples exposed each month. Compared with the millions of white samples, this part of the black samples is likely to be insufficient to reflect the full picture of the risk events. In other words, insufficient black samples will lead to insufficient robustness of the risk control trusted model.

[0030] Robustness is an important evaluation indicator of machine learning models. It is mainly used to test whether the model can still maintain the accuracy of judgment when facing small changes in input data, that is, whether the model's performance is stable when facing certain changes. The level of robustness directly determines the generalization ability of machine learning models.

[0031] In addition, in risk control scenarios, traditional neural network training methods do not meet the model interpretability requirements of trusted businesses. The training process of neural networks is a black box and lacks guidance. The final results are statistically available, but they may not meet the interpretability requirements when applied to individual cases. For example, a strong expert experience is that the longer the time since the first use of the user and the device (recency type features), the more trustworthy the device is. However, a black sample case of stolen data (R=30) in the data sample caused the final trained model to be distorted in this type of feature, resulting in a local abnormal interval.

[0032] Therefore, the present disclosure provides a fusion training scheme based on knowledge representation learning for a risk control model, which is based on the accumulation of expert experience in the field of risk control, and improves the robustness of the risk control model by incorporating multi-order feature intersection and data purification, and makes the interpretability of the risk control model meet the requirements.

[0033] In this disclosure, the electronic payment risk control will be mainly used as an example to describe the specific scheme. Those skilled in the art can understand that the fusion training scheme based on knowledge representation learning of the risk control model disclosed in this disclosure is applicable to various types of risk control models, and is not limited to the electronic payment risk control model.

[0034] Figure 1 is a flowchart showing a fusion training method 100 based on knowledge representation learning of a risk control model according to an embodiment of the present disclosure.

[0035] There are two conventional methods for risk control trust: based on trusted manual strategies and based on black and white sample training trusted models. In the method based on trusted manual strategies, the initial trusted release of the risk control system depends on manual strategies. For example, if the first use time between the user and the device is >30 days ago, the number of days used is >10 days, and the cumulative amount exceeds 200 yuan, it is determined to be a trusted device, and a trusted release is given based on whether the device has been stolen. Because it is based on manual rules, the granularity is relatively coarse, and the precision and recall are low.

[0036] In the method of training a trusted model based on black and white samples, a real-time model is used to score each transaction event, and the one with a high trusted model score is given a trusted release. However, due to insufficient black samples, the trusted model is not robust enough, and the traditional neural network training method cannot meet the model explainability requirements of trusted business.

[0037] Knowledge representation learning is used to learn the distributed representation of entities and relationships. It expresses entities and relationships with reasonable triples based on the selection of appropriate representation space, and uses corresponding encoding models to model relationship interactions. This disclosure introduces expert knowledge auxiliary training in the modeling process of the risk control trust model through knowledge representation learning for expert knowledge, so as to improve the robustness of the risk control trust model and make the risk control trust model interpretable.

[0038] At 102, label data is received and expert knowledge is extracted.

[0039] Step 102 is the preparation of training data, namely black and white label data and rule extraction based on expert knowledge. The preparation of black and white label data is the same as in the traditional sense: based on transaction event samples, black and white labels are added through user complaints and qualitative review results.

[0040] The role of expert knowledge is to assist in training and purify labeled data, reduce the part of the model's output that is inconsistent with the meaning precipitated in the risk control scenario, and enhance the robustness of the trusted model.

[0041] At 106 , multi-order feature crossover is performed on the label data and the expert knowledge to obtain data representation and rule representation.

[0042] For multi-dimensional input feature data sets, feature crossover can perform nonlinear feature fitting, thereby improving the model's ability to model nonlinearity and, in turn, improving the model's performance. In the risk control trust model, there are not only multi-dimensional features, but also the importance of various features is quite different, that is, these features are non-homogeneous, so introducing multi-order feature crossover for features of different types and dimensions is conducive to improving the performance of the risk control trust model.

[0043] Expert knowledge is introduced because there is a lot of expert experience in the field of risk control. In addition, there are problems of attack and defense in the field of risk control, and the pace of change of its samples is very fast, which makes the data distribution of the model fluctuate greatly. In the construction and training of the risk control trust model, adding expert knowledge instead of just using labeled data helps to purify the labeled data and remove the abnormal value fluctuations in a small number of black samples.

[0044] Therefore, in the present disclosure, multi-order feature intersection is also performed on the introduced expert knowledge in order to obtain both data representation and rule representation.

[0045] The multi-order feature crossover of label data and expert knowledge is realized through data encoder and rule encoder respectively. Multi-order feature crossover includes first-order feature crossover, second-order feature crossover and high-order feature crossover ( In different application scenarios, different orders of feature crossover can be used as needed.

[0046] In one embodiment of the present disclosure, the first-order feature crossover uses a multi-layer perceptron MLP, the second-order feature crossover uses a factorization machine FM, and the high-order feature crossover uses a logarithmic neural network LNN.

[0047] Those skilled in the art can understand that both the multi-layer perceptron MLP and the factorization machine FM can be applied to the first-order feature crossover or the second-order feature crossover, and the third-order and higher-order feature crossover can adopt the deep crossover DeepCross, the high-order factorization machine HOFM, the very deep factorization machine xDeepFM, the deformable convolution DCN-V2, etc. Further, the above mechanism is not limited, and the new feature crossover mechanism can also be incorporated into the technical solution of the present disclosure.

[0048] The following will be combined Figure 3 A feature cross-implementation framework in a risk control scenario according to an embodiment of the present disclosure is described in detail.

[0049] At 108 , the data representation is refined based on the rule representation.

[0050] As mentioned above, there are problems of attack and defense in the field of risk control, and the samples change very quickly, which makes the data distribution of the model fluctuate greatly. Therefore, it is necessary to purify the label data to effectively control the fluctuation of the model data distribution, so that the purified information is relatively static, thereby improving the robustness of the model.

[0051] In an embodiment of the present disclosure, purifying the data representation based on the rule representation further includes using a decision block composed of multiple expert blocks to purify the data representation. Using a decision block composed of multiple expert blocks to purify the data representation can be achieved through expert blocks with different weights.

[0052] Of course, those skilled in the art can understand that in different application scenarios, the number of expert blocks can be adopted as needed, and the weights of the expert blocks can also be set or changed as needed.

[0053] In another embodiment of the present disclosure, purifying the data representation based on the rule representation includes introducing a rule-related loss function based on the rule representation. Subsequently, a fusion loss function of the rule-related loss function and the task-related loss function is constructed.

[0054] The following will be combined Figure 5 The data purification and model training process in the fusion training based on knowledge representation learning of the risk control model according to one embodiment of the present disclosure is described in detail.

[0055] At 110 , a risk control model is trained and output based on the purified data representation.

[0056] In one embodiment of the present disclosure, training a risk control model based on the purified data representation includes optimizing a constructed fusion loss function. When the fusion loss function reaches an optimum, the trained risk control model is output. The trained risk control model can be put into operation online.

[0057] Therefore, the knowledge representation learning-based fusion training method of the risk control model disclosed in the present invention is based on the accumulation of expert experience in the field of risk control. It improves the robustness of the risk control model by incorporating multi-order feature intersection and data purification, and makes the interpretability of the risk control model meet the requirements.

[0058] Figure 2 is a schematic diagram showing a model fusion training framework based on knowledge representation learning according to an embodiment of the present disclosure.

[0059] like Figure 2 As shown, the present disclosure discloses a model fusion training framework DeepWIS (Deep Learning based Trust Recognition Architecture with the Mixture of Expert Wisdom) based on knowledge representation learning.

[0060] The DeepWIS framework disclosed in the present invention is based on DeepCTRL (deep neural network with controllable rule representation), but the rule encoder and data encoder adopt the HORN (High Order Networks) structure for multi-order feature crossover, and incorporates expert block combinations of different numbers or weights as decision blocks for different tasks.

[0061] Specifically, Figure 2As shown in FIG. 1 , the underlying features on which the DeepWIS framework of the present disclosure is based are the label data set and the refined expert knowledge. It is then processed by two encoders, namely the rule encoder and the data encoder in the form of HORN. The features are parallelly processed by two encoding layers for high-dimensional feature crossover, namely the first-order, second-order and high-order (i.e., third-order and above) feature crossover.

[0062] After encoding, the rule encoder and data encoder generate two representation vectors Z respectively. r (Rule Representation) and Z d (Data representation), the two are concatenated after weighting to form a vector z:

[0063]

[0064] in, The value of is not fixed, but changes during the training process from The distribution of is randomly sampled to improve the generalization performance of the model between label tasks and knowledge tasks.

[0065] In another embodiment of the present disclosure, The distribution can be used The concatenation operation can be replaced by a vector bit-by-bit addition operation.

[0066] The vector z is followed by a decision block, which can be in the form of a simple MLP or other forms, such as Mixture-of-Experts (MoE).

[0067] Then, the loss of expert knowledge is calculated separately (i.e., the rule-related loss function) and the loss for the risk task (i.e., the task-related loss function). The weights of the two are given by Variable adjustment:

[0068]

[0069] Expert knowledge loss or rule loss The introduction of makes the purified information relatively static, thereby improving the robustness of the model. For different risk control scenarios, The settings can be made as needed and are compatible with Fusion can also be done on demand.

[0070] In one embodiment of the present disclosure, to balance and The initial loss ratio will be calculated first. , and then construct the fusion loss function of the rule-related loss function and the task-related loss function:

[0071]

[0072] The risk control trust model is trained under the DeepWIS framework and optimized for the final weighted target L. After the training converges, the trained model file is produced for subsequent online scoring.

[0073] Figure 3 is a schematic diagram showing a multi-order feature crossover process of label data and expert knowledge for a risk control model according to an embodiment of the present disclosure.

[0074] like Figure 3 As shown, according to an embodiment of the present disclosure, the multi-order feature crossover process of the label data and expert knowledge for the risk control model is performed using a data encoder and a rule encoder in the form of HORN. The label data and expert knowledge are passed through two encoding layers in parallel. The data encoding layer and the rule encoding layer perform multi-order feature crossover.

[0075] In one embodiment of the present disclosure, the multi-level feature crossover includes an MLP layer, an FM layer, and an LNN layer, which are respectively 1, 2, The characteristic cross of the order is shown in the following formula:

[0076]

[0077] in Represents activation function, such as Relu, Sigmoid, etc.

[0078] Represents the MLP layer, performing first-order feature crossover, are the parameters of the MLP layer, Represents the output of the Embedding Layer concatenated together.

[0079] Represents the FM layer (Factorization Machine), which performs second-order feature crossover. d𝑒 represents the number of fields. Represents the parameters of FM. Represents the embedding output of the i-th Field in EmbeddingLayer.

[0080] It stands for LNN layer (Logarithmic Neural Network), which performs high-order feature crossover of the third order or above. , o represents the order of feature crossover, starting from 3. Represents the embedding output of the i-th Field in the Embedding Layer. are the parameters of LNN.

[0081] As mentioned above, those skilled in the art can understand that both the multi-layer perceptron MLP and the factorization machine FM can be applied to the first-order feature crossover or the second-order feature crossover, and the third-order and higher-order feature crossover can adopt the deep crossover DeepCross, the high-order factorization machine HOFM, the very deep factorization machine xDeepFM, the deformable convolution DCN-V2, etc. Similarly, the third-order and higher-order high-order feature crossover listed above can also be applied to the first-order feature crossover or the second-order feature crossover.

[0082] Furthermore, the above mechanism is not limited, and a new feature crossover mechanism can also be incorporated into the technical solution of the present disclosure. Those skilled in the art can understand that different crossover feature mechanisms can be used depending on the application scenario.

[0083] Figure 4 It is a schematic diagram showing a feature cross-implementation framework in a risk control scenario according to an embodiment of the present disclosure.

[0084] like Figure 4 As shown, in risk control trusted business scenarios, feature patterns are often divided into single-subject (for example, features in the active party dimension, features in the passive party dimension, etc.), dual-subject (for example, active party-passive party dimension, active party-device dimension, etc.), multi-subject (for example, active party-device-passive party, etc.), etc.

[0085] In the fusion training scheme based on knowledge representation learning of the risk control model disclosed in the present invention, based on the feature cross-implementation framework, manual feature engineering only needs to be designed to dual subjects at most, and high-order cross-implementation is automatically completed by the feature cross-implementation framework. For example, for the three features of "the active party's current transaction and its own proportion in the past 7 days, the average transaction amount of the active party and the passive party in the past 7 days, and the proportion of transactions complained by the passive party within 90 days" of the dual subjects, high-level semantics such as "the probability of a risk occurring in an account with a 3-fold mutation in the transaction amount on an unfamiliar account with a complaint ratio higher than 20% is high, and it should not be released as trusted" can be automatically fitted.

[0086] Compared with the application of ordinary multi-layer perceptron MLP, the application of models that distinguish the intersection of 1, 2, and 3+ order features has improved results. This is because the characteristics of risk control application scenarios are different from those of text, image, voice and other tasks in that the importance of their features is not averaged. For images, every pixel is homogeneous; while the features in the risk control field are non-homogeneous.

[0087] For example, in the risk control field, there are velocity features such as "the maximum model score of the passive party account in the identification model in the past 7 days", which have fast data growth, fast processing speed, and high timeliness requirements. In experience, such features can often achieve better results through first-order semantics, and higher-order intersections sometimes bury the semantics of the features.

[0088] For example, features such as "summary value of account transaction amount in the past 90 days" often require The characteristic crossover of the first order (i.e., 3+ order) can play a more important role.

[0089] Therefore, in the knowledge representation learning-based fusion training scheme of the risk control model disclosed in the present invention, the 1, 2, 3+ order feature crossovers adopted by the feature crossover implementation framework have a good harmonizing effect on various types of features.

[0090] Figure 5 It is a schematic diagram showing the data purification and model training process in the fusion training based on knowledge representation learning of the risk control model according to an embodiment of the present disclosure.

[0091] In the fusion training scheme based on knowledge representation learning of the risk control model disclosed in the present invention, data purification is implemented in the decision block. As mentioned above, the decision block can be in the form of a simple MLP or other forms, such as a mixed expert MoE.

[0092] In one embodiment of the present disclosure, the decision block is implemented as a multi-gate mixture of experts (MMoE) layer. The MMoE structure adopted by the multi-task network structure uses n expert modules to simulate n expert scoring, and controls the weight of each expert's scoring for each task through a threshold mechanism, as shown in the following formula:

[0093]

[0094] x is the output of the concatenation layer. k represents k tasks, and n represents n expert networks.

[0095] For different tasks, the output of a specific threshold represents the probability of different experts being selected, and multiple experts are weighted and summed. g(x) represents the output of the threshold (Gate), represents the weight of the k-th task on the upper threshold of the i-th expert, multiplied by the score of the i-th expert . It is a tower layer, which is used to obtain information unique to each task. It is generally a linear transformation plus a Softmax layer.

[0096] Represents the output of the Gate, which uses a multi-layer perceptron model and can be implemented as a linear transformation plus a Softmax layer.

[0097] Here, the difference from the traditional multi-task MoE structure is that the final y k The output is a single value, which degenerates into a multi-label task.

[0098] By introducing multiple experts into the decision-making block, the rapidly changing data features in the risk control field can be purified, and the outlier fluctuations in a small number of black samples will not have a decisive impact on the final model.

[0099] Furthermore, the loss function of the expert experience embodied by the rule encoder is introduced into the risk control trust model. .

[0100] The reason why traditional models are not robust is that black samples in the risk control field are not absolute. On the one hand, black samples are manually determined and have certain errors; on the other hand, cases audited by the existing system will become hidden cases. In addition, there are problems of attack and defense in the risk control field, which makes the pace of change of black samples very fast.

[0101] After adding the rules output by expert experience to improve the loss, this type of purified information is relatively static and can effectively control the data distribution of the model from fluctuating too much.

[0102]

[0103] and The weights between This is to guide the rule encoder and the data encoder to learn their corresponding semantics respectively.

[0104] Specifically, the loss of risky tasks It is no different from traditional tasks. When there are multiple risk tasks, a weighted summary is required first:

[0105]

[0106] As mentioned above, in one embodiment of the present disclosure, to balance and The initial loss ratio will be calculated first. , and then construct the fusion loss function of the rule-related loss function and the task-related loss function:

[0107]

[0108] For example, in the device trust semantics of determining whether a transaction involves the risk of theft, the time since the first successful transaction of an account and device has a strong positive correlation with the degree of trust, and there is a threshold of 7 days, which represents the average reporting cycle of users. However, during the model training process, it is easy to be affected by extreme values ​​of dirty data, resulting in fluctuations in the output of the model and inconsistent with expert cognition.

[0109] In a sense, we can extract a piece of expert knowledge that "if the device was first used more than 7 days ago than if it was used less than 7 days ago, the device is more credible in terms of misappropriation semantics."

[0110] The input feature vector is x , where the feature is x k . Introducing a small bias term ,but The output of the model before and after adding the bias is y j and y p,j , then the loss term of the knowledge represented by the above semantics is:

[0111]

[0112] Where a=7. The meaning of this formula is that when x k and x p If the model scores on both sides of the threshold 7 and the current device is first used for more than 7 days are less trustworthy than those for less than 7 days, a penalty will be imposed.

[0113] When there is multiple expert knowledge, weighted aggregation is required first:

[0114]

[0115] During model training, the final weighted target L is optimized. After training convergence, the trained model file is generated for subsequent online scoring.

[0116] Therefore, in the data purification and training process disclosed in the present invention, based on the accumulated experience of experts in the risk control field, the problem of model robustness in the risk control field is solved, while also meeting the demand for explainability in the risk control field.

[0117] Figure 6 is a block diagram showing a fusion training system 600 based on knowledge representation learning of a risk control model according to an embodiment of the present disclosure.

[0118] According to an embodiment of the present disclosure, a fusion training system 600 based on knowledge representation learning of a risk control model includes an information acquisition module 602 , an adjustment cross module 606 , a purification module 608 and a training module 610 .

[0119] The information acquisition module 602 receives label data and extracts expert knowledge.

[0120] The information acquisition module 602 prepares the training data, namely the black and white label data and the rule extraction based on expert knowledge. The black and white label data is prepared in the same way as in the traditional sense: based on the transaction event samples, the black and white labels are marked through the results of user complaints and trial qualitative assessment. The role of expert knowledge is to assist in training and purify the label data, reduce the part of the model output that does not match the meaning precipitated in the risk control scenario, and enhance the robustness of the credible model.

[0121] The feature crossover module 606 performs multi-order feature crossover on the label data and expert knowledge to obtain data representation and rule representation.

[0122] For multi-dimensional input feature data sets, feature crossover can perform nonlinear feature fitting, thereby improving the model's ability to model nonlinearity and, in turn, improving the model's performance. In the risk control trust model, there are not only multi-dimensional features, but also the importance of various features is quite different, that is, these features are non-homogeneous, so introducing multi-order feature crossover for features of different types and dimensions is conducive to improving the performance of the risk control trust model.

[0123] Expert knowledge is introduced because there is a lot of expert experience in the field of risk control. In addition, there are problems of attack and defense in the field of risk control, and the pace of change of its samples is very fast, which makes the data distribution of the model fluctuate greatly. In the construction and training of the risk control trust model, adding expert knowledge instead of just using labeled data helps to purify the labeled data and remove the abnormal value fluctuations in a small number of black samples.

[0124] Therefore, in the present disclosure, multi-order feature intersection is also performed on the introduced expert knowledge in order to obtain both data representation and rule representation.

[0125] The multi-order feature crossover of label data and expert knowledge is realized through data encoder and rule encoder respectively. Multi-order feature crossover includes first-order feature crossover, second-order feature crossover and high-order feature crossover ( In different application scenarios, different orders of feature crossover can be used as needed.

[0126] The purification module 608 purifies the data representation based on the rule representation.

[0127] As mentioned above, there are problems of attack and defense in the field of risk control, and the samples change very quickly, which makes the data distribution of the model fluctuate greatly. Therefore, it is necessary to purify the label data to effectively control the fluctuation of the model data distribution, so that the purified information is relatively static, thereby improving the robustness of the model.

[0128] The purification module 608 purifies the data representation based on the rule representation, which may include the purification module 608 using a decision block composed of multiple expert blocks to purify the data representation. Using a decision block composed of multiple expert blocks to purify the data representation may be achieved through expert blocks with different weights.

[0129] Of course, those skilled in the art can understand that in different application scenarios, the number of expert blocks can be adopted as needed, and the weights of the expert blocks can also be set or changed as needed.

[0130] The purification module 608 purifies the data representation based on the rule representation and may further include the purification module 608 introducing a rule-related loss function based on the rule representation. Subsequently, the purification module 608 constructs a fusion loss function of the rule-related loss function and the task-related loss function.

[0131] The training module 610 trains and outputs a risk control model based on the purified data representation.

[0132] The training module 610 trains the risk control model based on the purified data representation, including optimizing the constructed fusion loss function by the training module 610. When the fusion loss function reaches the optimum, the training module 610 outputs the trained risk control model. The trained risk control model can be put into operation online.

[0133] Therefore, the knowledge representation learning-based fusion training system of the risk control model disclosed in the present invention is based on the accumulated experience of experts in the field of risk control. It improves the robustness of the risk control model by incorporating multi-order feature intersection and data purification, and makes the interpretability of the risk control model meet the requirements.

[0134] The various steps and modules of the fusion training method and system based on knowledge representation learning of the risk control model described above can be implemented in hardware, software, or a combination thereof. If implemented in hardware, the various illustrative steps, modules, and circuits described in conjunction with the present invention can be implemented or executed with a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic components, hardware components, or any combination thereof. The general-purpose processor can be a processor, a microprocessor, a controller, a microcontroller, or a state machine, etc. If implemented in software, the various illustrative steps and modules described in conjunction with the present invention can be stored or transmitted as one or more instructions or codes on a computer-readable medium. The software modules that implement the various operations of the present invention may reside in a storage medium, such as RAM, flash memory, ROM, EPROM, EEPROM, registers, hard disks, removable disks, CD-ROMs, cloud storage, etc. The storage medium may be coupled to a processor so that the processor can read and write information from / to the storage medium, and execute the corresponding program modules to implement the various steps of the present invention. Moreover, the software-based embodiment can be uploaded, downloaded, or remotely accessed by appropriate communication means. Such appropriate communications means include, for example, the Internet, the World Wide Web, an intranet, software applications, cable (including fiber optic cables), magnetic communications, electromagnetic communications (including RF, microwave and infrared communications), electronic communications or other such communications means.

[0135] It should also be noted that these embodiments may be described as a process depicted as a flow chart, flow diagram, structure diagram, or block diagram. Although the flow chart may describe the operations as sequential processes, many of these operations can be performed in parallel or concurrently. In addition, the order of these operations can be rearranged.

[0136] The disclosed systems, devices and systems should not be limited in any way. On the contrary, the present invention covers all novel and non-obvious features and aspects of the various disclosed embodiments (individually and in various combinations and sub-combinations with each other). The disclosed systems, devices and systems are not limited to any specific aspects or features or combinations thereof, and any disclosed embodiment does not require the existence of any one or more specific advantages or the resolution of specific or all technical problems.

[0137] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many changes without departing from the scope of protection of the present invention and the claims, all of which fall within the scope of protection of the present invention.

Claims

1. A fusion training method for a risk control model, comprising: Receive labeled data and extract expert knowledge; Perform multi-order feature crossover on the label data and the expert knowledge to obtain data representation and rule representation; Purifying the data representation based on the rule representation; as well as The risk control model is trained and output based on the purified data characterization.

2. The method as claimed in claim 1, wherein the multi-order feature crossover includes a first-order feature crossover, a second-order feature crossover and a high-order feature crossover.

3. According to the method described in claim 1, the label data and the expert knowledge are subjected to multi-order feature crossover respectively through a data encoder and a rule encoder. 4 . The method of claim 1 , wherein refining the data representation based on the rule representation further comprises using a decision block composed of a plurality of expert blocks to purify the data representation.

5. The method of claim 1, wherein purifying the data representation based on the rule representation comprises introducing a rule-related loss function based on the rule representation, wherein the rule-related loss function comprises a loss of the expert knowledge.

6. The method as claimed in claim 5, purifying the data representation based on the rule representation also includes constructing a fusion loss function of the rule-related loss function and the task-related loss function, and the task-related loss function includes the loss of risky tasks. The method according to claim 1 , wherein the label data is black and white label data.

8. The method as claimed in claim 2, wherein the first-order feature crossover adopts a multi-layer perceptron MLP, the second-order feature crossover adopts a factorization machine FM, and the high-order feature crossover adopts a logarithmic neural network LNN.

9. The method according to claim 4, wherein the purifying of the data representation by using a decision block composed of a plurality of expert blocks is achieved by expert blocks with different weights.

10. The method as claimed in claim 6, wherein training the risk control model based on the purified data representation includes optimizing the constructed fusion loss function.

11. A fusion training system for a risk control model, comprising: Information acquisition module, which receives label data and extracts expert knowledge; A feature crossover module performs multi-order feature crossover on the label data and the expert knowledge to obtain data representation and rule representation; A purification module, purifying the data representation based on the rule representation; as well as The training module trains and outputs the risk control model based on the purified data representation.

12. A computer-readable storage medium storing instructions, which, when executed, causes a machine to perform the method according to any one of claims 1 to 10.

Images