A method for predicting online loan fraud based on multi-stage data representation

Abstract

The present invention belongs to the technical field of online loan fraud prediction, and discloses an online loan fraud prediction method based on multi-stage data representation, comprising the following steps: step S101: select available fields according to original online loan application data to construct an original information network, then divide the nodes into entities and attributes, and construct a hierarchical information network consisting of an entity layer and an attribute layer; step S102: extract an attribute network from the attribute layer, attach a pre-trained word vector to each attribute to reflect the semantic similarity between the attributes, and use network representation learning to integrate the attribute similarity information and attribute co-occurrence association obtained from prior and external semantic knowledge into the embedding vector corresponding to the attribute; step S103, entity representation; step S104, fraud prediction. The present invention enhances the density of information association in online loan application transactions, improves the detection capability of online loan fraud prediction models, and provides better protection for the financial security of users and enterprises.

Description

Technical Field

[0001] The present invention relates to the technical field of online loan fraud prediction, and in particular to an online loan fraud prediction method based on multi-stage data representation. Background Art

[0002] At present, with the increasing expansion of Internet business, lending has also entered a new stage of development. The Internet has brought convenience to lending, lowered the threshold for lending, and achieved a large increase in micro loans. However, the complexity of the Internet environment and the low threshold for online lending increase the possibility of fraud in online lending. Fraudulent behavior in online lending often brings uncontrollable huge losses to enterprises and society, causing structural risks. In order to ensure the business security of online lending systems in Internet scenarios, it is necessary to establish an effective online lending fraud prediction method.

[0003] Traditional online lending anti-fraud technologies such as blacklists and whitelists and business rules, as the first line of defense for online lending anti-fraud security, are not only difficult to cope with the rapid evolution of fraud technology in the Internet scenario, but also insufficient to fully cover the entire target user group of online lending. Therefore, machine learning anti-fraud technology based on the digital-driven paradigm can serve as the second line of defense to supplement the field of online lending anti-fraud security protection.

[0004] At present, the data-driven paradigm is an effective solution for digital financial anti-fraud tasks, and related work has been widely studied. It has become an industry consensus that machine learning algorithms can help achieve better anti-fraud in online lending. For example, Liang et al. proposed an automatic fraud detection solution based on graph learning algorithms at the ACM SIGIR 2019 conference to separate fraudsters from old customers and discover organized fraudsters. In addition, some researchers have further successfully combined machine learning technology and network representation learning technology, such as graph neural networks, to mine deep associations hidden in applicant data. However, the current scale of online lending data is often large, but fraud data is hidden in a large amount of normal data, which brings huge challenges to online lending anti-fraud prediction. Because the data is limited, useful information for fraud prediction is usually insufficient, and it is difficult to establish a data-driven paradigm model based solely on application information with low information density.

[0005] Through the above research, we found that a major problem in online loan fraud prediction can be attributed to the fact that the information density of the application information is too low to be directly applied to the online loan fraud prediction task. Some existing related works directly operate on the carrier of complete original information to fully preserve useful information, but the effect is usually poor because the useful information is buried in a large amount of useless information. Another part of the work reduces the size of the information carrier by discarding part of the information to overcome the influence of useless information, which leads to incomplete information and usually greatly reduces the performance of fraud prediction. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a multi-stage data characterization method based on network representation learning as an enhancer for improving information density. At the same time, the information density is improved by increasing useful information and reducing information carriers, thereby achieving data enhancement and improving the anti-fraud ability of the machine learning model.

[0007] The technical problem to be solved by the present invention is that the traditional fraud prediction method in the online loan anti-fraud scenario is difficult to deal with sparsely associated loan data and cannot capture the deep correlation between loan applications. Therefore, it is difficult to build an efficient anti-fraud model using only basic application information, and it lacks good generalization ability.

[0008] In order to solve the above technical problems, the technical solution of the present invention is:

[0009] A method for predicting online loan fraud based on multi-stage data representation, characterized by being divided into four steps:

[0010] Step S101, association representation:

[0011] Select available fields according to the field missing rate in the original online loan application data, build an original information network based on the fields, and then divide the nodes in the original information network into entities and attributes to build a hierarchical information network consisting of an entity layer and an attribute layer; provide it to S102 and S103;

[0012] Among them, the missing rate refers to the ratio of the number of times a field is missing in the original online lending dataset to the total number of lending transactions;

[0013] Step S102, attribute characterization:

[0014] Extracting an attribute network from the attribute layer in the hierarchical information network, attaching a pre-trained word vector to each attribute to reflect the semantic similarity between the attributes, and integrating the attribute similarity information and attribute co-occurrence association obtained from prior and external semantic knowledge into the embedding vector corresponding to the attribute by using network representation learning; providing to S103;

[0015] Step S103, entity representation:

[0016] Extracting an entity network from the entity layer of the hierarchical information network, generating an entity network with attributes by using the layer correspondence association and the embedding vector obtained in step S102, and embedding the entity network with attributes into a low-dimensional vector space by using network representation learning to obtain an embedding vector of the entity; providing it to S104;

[0017] Step S104, fraud prediction:

[0018] The similarity between an entity and its first-order neighbors (entities) is introduced as a micro-feature in an entity network with attributes. The similarity between loan application transactions and their second-order neighbors in the entity network is used as a macro-feature. The similarity of the embedding vectors of different entities is fed into a common classifier model to realize fraud prediction.

[0019] The process of step S101 is as follows:

[0020] Obtaining original online loan application data, and setting a training set time window and a test set time window based on test requirements, so as to divide the original online loan application data into training set data and test set data, so that the training set data and the test set data constitute online loan data that can be used for subsequent models;

[0021] Perform field selection on the original online lending data;

[0022] The fields of the selected initial online loan data are extracted for association, and the association between the online loan data is described using a graph structure. For the selected fields of an application in the original online loan data, an original information network is constructed using a star topology structure: its identifier (i.e., the number that uniquely indicates a loan application in the data set, which generally has no practical meaning, such as the order number of a loan order) and the values ​​of all fields (such as the applicant, the applicant's address, the applicant's mobile phone number, etc. in a loan application) are represented as nodes, and the identifier and the corresponding field value are respectively recorded as two nodes constituting an edge, and different loan application identifiers are connected through their shared fields;

[0023] Dividing nodes in the original information network into entities and attributes;

[0024] The entity set V is divided into V t and V d There are two parts. For each loan application, its identifier is recorded as v t ∈V t , and denote the other entities in the application as v d ∈V d Entity v d and the identifier vt The co-occurrence relationship (i.e., co-appearance in an application) is recorded as e = (v t , v d , r)∈E (E is the set of all co-occurrence associations), where type r∈R (R is the set of descriptions of all entities in the original data) is the type of entity v in the original data. d Description, collection (× is the Cartesian product operation);

[0025] All attributes θ v The set Θ of is recorded as the attribute layer in the hierarchical information network;

[0026] Extract the corresponding association between entities and attributes from the loan application (i.e., an attribute describes an entity correspondingly). In the original information network, attribute θ v and entity v d Both are connected to entity v t , the attribute θ v and entity v d The common connection between them is recorded as the layer corresponding association (v d ,θ v ), that is, the corresponding association between entities and attributes;

[0027] In the attribute layer, the corresponding attribute of each entity v∈V is recorded as Wherein i represents the type of attribute. By processing the original information network, the entity layer, the attribute layer and the corresponding association of the layers can be obtained, and then a hierarchical information network is constructed by the entity layer and the attribute layer and the corresponding association of the layers.

[0028] Preferably, fields with a missing rate lower than 50% are selected as available fields.

[0029] Preferably, the nodes in the original information network are divided into entities and attributes according to the principles of uniqueness and indivisibility;

[0030] The uniqueness means that the selected field can uniquely represent the same object among multiple fields describing an object;

[0031] The indivisibility means that an entity does not contain potential information, that is, a unique value can be selected as an entity, and then the entity has no overlapping meaning with other entities;

[0032] Fields that meet the requirements of uniqueness and indivisibility are divided into entities, and the set V of entities is recorded as the entity layer in the hierarchical information network;

[0033] After the entity is selected, the remaining fields are divided into attributes, and all attributes θ v The set Θ of is recorded as the attribute layer in the hierarchical information network.

[0034] The process of step S102, attribute characterization, is as follows:

[0035] Extract an attribute network G′ from the attribute layer in the hierarchical information network Θ =(Θ,E Θ , R Θ ) to reflect the co-occurrence association between attributes, the co-occurrence of two attributes corresponding to the same entity in the entity layer is recorded as an edge in the attribute network, the number of co-occurrences is used as the weight ω of the association between the attributes, and the edge is recorded as where r′∈R Θ Depends on two properties and Type,

[0036] A pre-trained word vector is attached to each attribute to reflect the semantic similarity between attributes. Furthermore, the pre-trained word vector is used as the vectorization of the attribute and the pre-trained word vector is recorded as Then, Denoted as a semantic attribute network, where is a word vector set;

[0037] Using network representation learning to learn semantic attribute network G Θ The embedding vector of the node in is based on the embedding vector, and each attribute θ v It can be represented as a vector

[0038] The process of step S103, entity characterization, is as follows:

[0039] Extract the entity network G from the entity layer of the layered information network V =(V, E, R), based on the attribute embedding vector described in step S102, each attribute θ v Convert to the corresponding vector Then, the attribute vector of each entity is connected by combining the entity network and the layer correspondence between the attribute and entity layers Furthermore, the entity network with attributes is recorded as in is the attribute θ v The corresponding vector;

[0040] In the entity network G with attributes V Apply the network representation learning algorithm to obtain the entity representation corresponding to each entity v in All associations in the data are preserved.

[0041] The process of step S104, fraud prediction, is as follows:

[0042] First, two automatic embedding features are designed to reflect the association between loan applications. The automatic embedding features are: observing the similarity between the application identifier and each connected entity from a micro perspective (association within the same loan application), and observing the similarity between the application identifier and other indirectly connected identifiers from a macro perspective (association between different loan applications).

[0043] Introduce entity v into the entity network with attributes t and its first-order neighbors (entity v d , where the first-order neighbors specifically refer to the entity v in the attributed entity network. t The similarity of nodes directly connected by edges is used as a micro-feature for entity v t The corresponding embedding vector and entity v d The corresponding embedding vector Calculating Similarity

[0044] Further, using the entity network v t The second-order neighbor (entity v′ t , where the second-order neighbors specifically refer to the entity v in the attributed entity network. t The similarity between nodes with common first-order neighbors is used as a macro feature. Given the entity v t The corresponding embedding vector and entity v′ t The corresponding embedding vector Calculating Similarity

[0045] Based on prior knowledge, some manual statistical features are designed, such as the proportion of fraudulent samples to all samples, the number of applications to be tested that are associated with fraudulent applications, etc.

[0046] Then, the automatic embedding features (i.e., the similarity of the embedding vectors of different entities) and the manual statistical features are fed into a commonly used classifier to fit the training samples to obtain a fraud prediction model. For the input test sample, after the feature generation, it is input into the fraud prediction model for judgment to obtain the probability that the test sample is fraudulent.

[0047] In particular, the missing similarity in the micro-features is replaced by a constant term, and only the first s features with the smallest similarity are retained as macro-features, where s is a hyperparameter that can be set to any integer in the interval [1,10].

[0048] Preferably, the similarity is calculated using Euclidean distance;

[0049] For entity v t and v d The embedding vector of , where D is the dimension, for The value of the i-th dimension of for The i-th dimension value of is calculated using the following Euclidean distance formula:

[0050]

[0051] For the embedding vector and for The similarity can be obtained by using the Euclidean distance calculation formula.

[0052]

[0053] Preferably, as described in S102, the attribute network G with semantics Θ and the entity network G with attributes described in S103 V Has a similar form, where the attribute set Θ (or the pre-trained word vector set G Θ ) and entity set V (or attribute embedding vector set ) plays the same role in the network representation learning algorithm. Therefore, the same network representation learning algorithm can be used in the attribute representation and entity representation stages to integrate different types of information into the corresponding embedding vectors.

[0054] Beneficial effects:

[0055] The online loan fraud prediction method based on multi-stage data representation provided by the present invention is applied to generate a hierarchical information network including an entity layer and an attribute layer by dividing all nodes in the original information network; then, for the attribute layer in the hierarchical information network, a pre-trained word vector is introduced to capture the semantic similarity between the attributes, and the attribute layer and the word vector are embedded into the vector space using a network representation learning algorithm to form an embedding vector corresponding to the attribute node, thereby realizing the introduction of external prior knowledge; further, for the entity layer in the hierarchical information network, the present invention combines the embedding vector corresponding to the attribute, embeds the entity layer and the embedding vector corresponding to the attribute into the vector space using a network representation learning algorithm to form an embedding vector corresponding to the entity node, thereby realizing the reduction of the network scale; the above steps gradually improve the information density of the online loan data; further, for the embedding vector corresponding to the obtained entity, features are constructed from a macroscopic and microscopic perspective respectively, and the macroscopic and microscopic features are input into a binary classification model for training and testing to obtain the fraud probability of the online loan data; compared with the traditional method, the present invention gradually improves the information density of the online loan data, and at the same time characterizes the fraud features from a macroscopic and microscopic perspective, with higher accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0057] Figure 1 A schematic diagram of the process of an online loan fraud prediction method based on multi-stage data representation according to an embodiment of the present invention is shown;

[0058] Figure 2 A schematic diagram of the working process of the online loan fraud detection system according to an embodiment of the present invention is shown;

[0059] Figure 3 The workflow of the online loan fraud prediction system of the online loan fraud prediction method based on multi-stage data representation in an embodiment of the present invention is shown;

[0060] Figure 4 A schematic diagram of multi-stage data representation of an online loan fraud prediction method based on multi-stage data representation according to an embodiment of the present invention is shown;

[0061] Figure 5 A schematic diagram of a process for generating a semantically-based attribute network in an online loan fraud prediction method based on multi-stage data representation according to an embodiment of the present invention is shown.

[0062] Figure 6A KS curve comparison diagram of the online loan fraud prediction method based on multi-stage data representation according to an embodiment of the present invention and a baseline anti-fraud model is shown. DETAILED DESCRIPTION

[0063] The following will describe the implementation methods of the present invention in detail with reference to the accompanying drawings and examples, so that the implementation process of how the present invention applies the network representation learning algorithm to solve practical problems and achieve the technical effects can be fully understood and implemented accordingly.

[0064] Online lending is widely used in e-commerce, e-government, and electronic business process of companies. The anomalies in online lending mainly include intermediary agency, gang crime, machine behavior, account theft, identity fraud, and serial transactions. Among them, identity fraud is a common online lending fraud, which refers to users falsifying the personal identity, property certificates and other materials they provide, and even using illegal means such as deception to obtain other people's information, and then impersonating others to defraud loans.

[0065] The main purpose of fraud prediction in online lending is to prevent bad debts. Bad debts refer to accounts receivable that the lender cannot recover from the borrower. Bad debts in online lending usually include two situations: one is that the borrower is unable to repay the loan due due to certain factors after the loan; the other is the borrower's subjective fraud, that is, deliberately preparing to deny the loan. Setting up a fraud prediction program based on the basic information of the applicant plays an important role in improving the efficiency of loan approval. In addition, for other links of loan approval, such as "credit assessment" and "credit review", it is usually based on the assumption that the applicant is not a fraudster and the application information is authentic and reliable. Therefore, the prediction and identification of fraudulent applications is a necessary procedure before other operations.

[0066] The traditional business rule logic prediction method has a slow review process, consumes a lot of manpower and material resources, is difficult to adapt to new and changing fraud methods, and lacks good generalization capabilities. General machine learning anti-fraud methods based on data paradigms cannot effectively mine information for low-quality applicant data and perform poorly in online lending scenarios.

[0067] Example:

[0068] In order to solve the problems existing in the prior art, an embodiment of the present invention provides an online loan fraud prediction method based on multi-stage data representation.

[0069] Figure 3 The workflow of the online loan fraud prediction method based on multi-stage data representation according to an embodiment of the present invention is shown; Figure 5 A schematic diagram of multi-stage data representation of an online loan fraud prediction method based on multi-stage data representation according to an embodiment of the present invention is shown; Figure 3 and Figure 4 As shown in FIG, the online loan fraud prediction method based on multi-stage data representation includes the following four steps: association representation, attribute representation, entity representation and fraud prediction.

[0070] Step S101, association characterization, select available fields according to the field missing rate in the original online loan application data (as shown in Table 1), build an original information network based on the selected fields, and then divide the nodes in the original information network into entities and attributes according to the proposed uniqueness and indivisibility principles, and build a hierarchical information network consisting of an entity layer and an attribute layer.

[0071] Table 1 Examples of common fields in original online loan application data

[0072]

[0073] Specifically, the original online loan application data is obtained, and the training set time window and the test set time window are set based on the test requirements to divide the training set data and the test set data from the original online loan application data. The training set time window and the test set time window are both continuous time windows, and in order to avoid the time travel problem, it should be ensured that the occurrence time of the original online loan application data in all training set data is earlier than the occurrence time of the online loan data in any test set data. Preferably, the training set time window is generally set according to the data scale. The larger the data scale, the smaller the training set window and the test set window should be. After the training set time window and the test set time window are set, the training set data is divided from the original online loan data based on the training set time window, and the test set data is divided from the original online loan data based on the test set time window, and the training set data and the test set data constitute the online loan data that can be used for subsequent models.

[0074] Then, the original online lending data is subjected to field selection. In the original online lending data set, each loan application transaction may consist of dozens of fields. However, due to the difficulty of data collection and the user's privacy policy, most fields have missing values. Too high a missing rate seriously affects the performance of the prediction model. Preferably, this embodiment only selects fields with a missing rate of less than 50% as available data. The missing rate refers to the ratio of the number of times a field is missing in all loan transactions in the original online lending data set to the total number of loan transactions.

[0075] The fields of the selected initial online loan data are associated and extracted. In order to better represent the associations between online loan data, this embodiment uses a graph structure to describe the associations between online loan data. For the fields of an application in the original online loan data (usually including the identity, work unit, social relationship information, etc. of the applicant), a star topology is used to construct an original information network, and its identifier and the values ​​of all fields are represented as nodes, and the identifier and the corresponding field value pairs are recorded as edges, and then different loan application identifiers are connected through their shared fields. In the original information network, redundancy between certain fields will cause the network scale to be too large and bury useful information. Some original fields are supplementary to other fields, and they should not be used as entities alone, such as an address and a city have overlapping meanings. Preferably, a part of the fields are selected as entities to represent the original application data, such as relatives, work units, telephones, emails, addresses, etc.

[0076] In the original information network, the determination of entities depends on two principles: uniqueness and indivisibility. Uniqueness means that the selected field can uniquely represent the same object among multiple fields describing an object. Indivisibility means that the entity does not contain potential information, that is, a unique value can be selected as an entity, then the entity has no overlapping meaning with other entities. For example, a detailed address uniquely represents a location and does not overlap with other detailed addresses, while the province information is only a coarse-grained description corresponding to multiple detailed addresses.

[0077] All entities that meet the uniqueness and indivisibility requirements are selected, and the entity set V is recorded as the entity layer in the hierarchical information network. In the entity layer, the entity set V is divided into two parts (V t and V d ). For each loan application, its identifier is recorded as v t ∈V t , and denote the other entities in the application as v d ∈V d . Then, entity v d and the identifier v t The co-occurrence relationship (i.e., co-appearance in an application) is recorded as e = (v t , v d , r)∈E, where type r is the entity v in the original data d Description, collection For example, when entity v d When it is a telephone number, the relationship r means that the applicant is applying for v t The phone number provided is v d .

[0078] Furthermore, the remaining fields after the entity is selected are taken as attributes, and all attributes θ v The set θ of is recorded as the attribute layer in the hierarchical information network. In the loan application, there are multiple fields describing the same entity. Then, the corresponding association between the entity and the attribute is extracted from the loan application (that is, the attribute describes an entity correspondingly). In the original information network, the attribute θ v and entity v d Both are connected to entity v t , the attribute θ v and entity v d The common connection between them is recorded as the layer corresponding association (v d ,θ v ), that is, the corresponding association between entities and attributes.

[0079] In the attribute layer, the corresponding attribute of each entity v∈V is recorded as Where i represents the type of attribute. By processing the original information network, the entity layer, the attribute layer and the layer corresponding association can be obtained. Then, a hierarchical information network is constructed by the entity layer and the attribute layer and the layer corresponding association.

[0080] Step S102: attribute representation, integrating attribute similarity information and attribute co-occurrence association obtained from prior and external semantic knowledge into the embedding vector corresponding to the attribute.

[0081] Specifically, firstly, an attribute network G′ is extracted from the attribute layer in the hierarchical information network. Θ =(Θ,E Θ , R Θ ) to reflect the co-occurrence association between attributes. When the two attributes correspond to the same entity in the entity layer, this co-occurrence is recorded as an edge in the attribute network. Different from the co-occurrence association between entities in the entity layer, the co-occurrence association between attributes may appear multiple times, that is, two attributes appear in multiple applications. Further, the number of co-occurrences is calculated as the weight ω of the association between attributes. The association between attributes is recorded as where r′∈R Θ Depends on two properties and Type, For example, the association between a city and a district should be different from the association between a city and a province.

[0082] In addition to the co-occurrence associations between attributes, there are similarities between the rich semantic information contained in the attributes. Furthermore, a pre-trained word embedding is attached to each attribute to reflect the semantic similarity between the attributes. Preferably, the pre-trained word embedding can be obtained by learning some prior and external semantic knowledge in data similar to lending scenarios through a natural language processing algorithm. In the case of limited data, public word embeddings trained based on general data (such as Google's word2vec, Facebook's fastText, and Stanford's GloVe) can also be used.

[0083] Furthermore, we use the pre-trained word vector as the vectorization of the attribute, and record the pre-trained word vector as Then, Denoted as a semantic attribute network, where is a word vector set. Network representation learning is a commonly used tool that can automatically capture deep hidden associations and map the network to a low-dimensional space. It uses network representation learning to learn semantic attribute networks G Θ Further, based on the embedding vector, each attribute θ v It can be represented as a vector The generation process of attribute networks with semantics is as follows Figure 5 shown.

[0084] Step S103, entity representation, using the layer correspondence association and the embedding vector obtained in step S102, to generate an entity network with attributes, and using network representation learning to embed the entity network with attributes into a low-dimensional vector space to obtain an embedding vector of the entity.

[0085] Specifically, firstly, the entity network G is extracted from the entity layer of the hierarchical information network. V =(V, E, R), to reflect the co-occurrence relationship between entities. Based on the attribute embedding vector described in step S102, each attribute θ v Convert to the corresponding vector The attribute embedding vector integrates the co-occurrence relationship and similarity relationship between attributes in step S102. Then, the attribute vector of each entity is connected by combining the entity network and the layer correspondence between the attribute and the entity layer. Furthermore, the entity network with attributes is recorded as in is the attribute θ v The corresponding vector. The relationship e∈E reflects the co-occurrence association between entities, and the similarity between the embedding vectors of attributes reflects the co-occurrence association and similarity association between attributes.

[0086] Similar to the attribute representation stage, in the process of learning the entity embedding vector, it is necessary to consider the co-occurrence association between entities and the similarity association between the attribute embedding vectors. V Apply the network representation learning algorithm to obtain the entity representation corresponding to each entity v in All the associations in the data are preserved. Preferably, the attribute network G with semantics Θ and the entity network G with attributes V Has a similar form, where the attribute set Θ (or the pre-trained word vector set G Θ ) and entity set V (or attribute embedding vector set ) plays the same role in the network representation learning algorithm. Therefore, the same network representation learning algorithm can be used in both the attribute representation and entity representation stages to integrate different types of information into the corresponding embedding vectors.

[0087] Step S104, fraud prediction. Inconsistency is the key to online loan fraud prediction, such as inconsistent node neighbors, inconsistent features, and inconsistent associations in the native network. Automatic embedding features and manual statistical features are designed to discover the above inconsistencies to achieve fraud prediction, and then predict whether an online loan application is a fraud application.

[0088] Specifically, the embedding vector of the entity can be applied to scenarios such as node classification, link prediction and community detection to achieve fraud prediction. Preferably, this embodiment designs to feed the similarity of the embedding vectors of different entities into a common classifier model to achieve fraud prediction function. Two automatic embedding features are designed to reflect the association between loan applications, and the automatic embedding features are: From a micro perspective (association within the same loan application), observe the similarity between the application identifier and each connected entity, and from a macro perspective (association between different loan applications), observe the similarity between the application identifier and other indirectly connected identifiers. The micro similarity reflects whether a loan application is reasonable, for example, whether an entity is false information. The macro similarity reveals the relative relationship between a loan application and other applications, such as whether it involves fraudsters falsifying data in batches. This embodiment uses Euclidean distance as a similarity calculation method between the embedding vectors of two entities. Preferably, Euclidean distance is only a feasible similarity calculation method, and other similarity calculation methods can also be used, with only performance differences. The introduction of entity v in the entity network with attributes described in this embodiment t and its first-order neighbors (entity v d ) as microscopic features. t and v dThe embedding vector of , where D is the dimension, for The value of the i-th dimension of for The i-th dimension value of is calculated using the following Euclidean distance formula:

[0089]

[0090] Furthermore, this embodiment adopts the loan application transaction v t and belonging to the entity network v t The similarity between the second-order neighbors of is used as a macro feature. Given the embedding vector and for The similarity can be obtained by using the Euclidean distance calculation formula.

[0091] In particular, loan applications have different numbers of first-order and second-order neighbors in the entity network. For micro features, the absence of an entity will result in a reduction in the number of first-order neighbors, so this embodiment replaces the missing similarity with a constant term. For macro features, the number of second-order neighbors is uncertain, so this embodiment only retains the first s features with the smallest similarity. Preferably, most loan applications are independent, so there is no need to set a large s.

[0092] Preferably, based on prior knowledge, some manual statistical features can be designed, such as the proportion of fraudulent samples to all samples, the number of applications to be detected that are associated with fraudulent applications, etc.

[0093] Then, this embodiment feeds the automatic embedding features and manual statistical features into XGBoost (a commonly used classifier) ​​to fit the training samples to obtain a fraud prediction model. For the input test sample, after the feature generation, it is input into the fraud prediction model for judgment to obtain the probability that the test sample is fraudulent.

[0094] The online loan data deployed in the verification experiment of this embodiment comes from a real online loan anonymized data set provided by an online loan platform, which contains online loan application data from January to July 2017. This embodiment uses the data set from January to June 2017 as training samples, and sets the data set from July 2017 as test samples. In the online loan scenario, when the application expires, the online loan platform can obtain the status of whether the user has successfully repaid the loan.

[0095] The experiment was conducted under Windows Server 2012 operating system, with 40 Intel Xeon E5-2640v4 CPUs and 128GB RAM. In general, according to the demand for online loan fraud prediction in real business, the KS value is a widely used reliable metric. Therefore, this embodiment uses the KS value to measure performance.

[0096] The detailed calculation process of KS value is as follows:

[0097] Given a test set, after testing the test set using the anti-fraud model, the model will output the fraud risk probability value of each sample. The probability values ​​corresponding to each sample are arranged in positive order and recorded as Select As the fraud classification threshold, calculate each group of thresholds as The corresponding TPR value and FPR value Then we can get the TPR value and FPR value set under different thresholds, which are recorded as

[0098]

[0099]

[0100] Then the KS value of the anti-fraud model is the maximum value of the difference between TPR and FPR under different thresholds, that is

[0101]

[0102] The performance comparison of the proposed embodiment and the baseline method for fraud prediction task is given to highlight the effectiveness of the embodiment on the online loan application dataset. The baseline methods compared include DT [1] ET-RNN [2] DDFP [3] 、 crDNN [4] 、semiGNN [5] And manual feature engineering method (Manual). The so-called manual feature engineering method refers to the features designed by experts for fraud prediction. The KS curve of the experimental results is as follows Figure 6 As shown, in this embodiment Figure 6The present embodiment achieves better performance than other methods, with the best KS value exceeding 0.46. The performance of semiGNN is second only to the present embodiment. Compared with the manual feature engineering method, the performance of crDNN is similar to that of the manual feature engineering method, but both are lower than the above methods. There is a certain gap between ET-RNN and the previous methods. DT has the worst performance, which is about 0.06 lower than the best KS value. The above experiments prove that the performance of the present embodiment can meet the requirements of the anti-fraud task and is better than the existing baseline method.

[0103] The present invention aims to overcome the difficulty that traditional online loan anti-fraud prediction cannot be applied to online loan fraud prediction tasks when the information density of application information is too low. A multi-stage data representation method based on network representation learning is proposed as an enhancer to improve information density. At the same time, the information density of associations is improved by increasing useful information and reducing associations as information carriers, thereby achieving data enhancement and improving the anti-fraud ability of machine learning models.

[0104] Although the embodiments disclosed in the present invention are as above, the contents described are only embodiments adopted for facilitating the understanding of the present invention and are not intended to limit the present invention. Any technician in the technical field to which the present invention belongs can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in the present invention, but the protection scope of the present invention shall still be subject to the scope defined in the attached claims.

[0105] References:

[0106] [1] Zeynep Boz, Dilek S.Ilker Birbil,et al.Reassessment and monitoring of loan applications with machine learning[J].Applied ArtificialIntelligence,2018,32(9-10):939–955.

[0107] [2]Beutel A,Akoglu L,Faloutsos C.Graph-Based User Behavior Modeling:From Prediction to Fraud Detection[C] / / Acm Sigkdd International Conference onKnowledge Discovery&Data Mining.ACM,2015: 2309-2310.

[0108] [3]Dawei Cheng,Yiyi Zhang,Fangzhou Yang,et al.A dynamic defaultprediction framework for networked-guarantee loans[C] / / ACM CIKM 2019,2019:2547–2555.

[0109] [4]Tan F,Hou X,Zhang J,et al.A Deep Learning Approach to CompetingRisKS Representation in Peer-to-Peer Lending[J].IEEE Transactions on NeuralNetworks and Learning Systems,2018,PP:1-10.

[0110] [5]Daixin Wang,Yuan Qi,Jianbin Lin,et al.A semi-supervised graphattentive network for financial fraud detection[C] / / IEEE ICDM 2019,2019:598–607

Claims

1. A method for predicting online loan fraud based on multi-stage data representation, characterized in that: There are four steps: Step S101, association representation: Select available fields according to the field missing rate in the original online loan application data, build an original information network based on the fields, and then divide the nodes in the original information network into entities and attributes to build a hierarchical information network consisting of an entity layer and an attribute layer; provide it to S102 and S103; Among them, the missing rate refers to the ratio of the number of times a field is missing in the original online lending dataset to the total number of lending transactions; Step S102, attribute characterization: Extracting an attribute network from the attribute layer in the hierarchical information network, attaching a pre-trained word vector to each attribute to reflect the semantic similarity between the attributes, and integrating the attribute similarity information and attribute co-occurrence association obtained from prior and external semantic knowledge into the embedding vector corresponding to the attribute by using network representation learning; providing to S103; Step S103, entity representation: Extracting an entity network from the entity layer of the hierarchical information network, generating an entity network with attributes by using the layer correspondence association and the embedding vector obtained in step S102, and embedding the entity network with attributes into a low-dimensional vector space by using network representation learning to obtain an embedding vector of the entity; providing it to S104; Step S104, fraud prediction: In the attributed entity network, the similarity between the entity and its first-order neighbors is introduced as a micro-feature. The similarity between the loan application transaction and the second-order neighbors in the entity network is used as a macro-feature. The similarity of the embedding vectors of different entities is fed into a common classifier model to achieve fraud prediction function. The process of step S101 is as follows: Obtaining original online loan application data, and setting a training set time window and a test set time window based on test requirements, so as to divide the original online loan application data into training set data and test set data, so that the training set data and the test set data constitute online loan data that can be used for subsequent models; Perform field selection on the original online lending data; The fields of the selected initial online loan data are extracted, and the association between the online loan data is described by using a graph structure. For the selected fields of an application in the original online loan data, an original information network is constructed using a star topology structure: its identifier and the values ​​of all fields are represented as nodes, and the identifier and the corresponding field value are respectively recorded as two nodes constituting an edge, and different loan application identifiers are connected through their shared fields; Dividing nodes in the original information network into entities and attributes; The entity set V is divided into V t and V d There are two parts. For each loan application, its identifier is recorded as v t ∈V t , and denote the other entities in the application as v d ∈V d Entity v d and the identifier v t The co-occurrence relationship is recorded as e = (v t , v d , r)∈E, where type r∈R, R is the set of descriptions of all entities in the original data, and is the entity v in the original data. d Description, collection All attributes θ v The set Θ of is recorded as the attribute layer in the hierarchical information network; Extract the corresponding association between entities and attributes from the loan application. In the original information network, attribute θ v and entity v d Both are connected to entity v t , the attribute θ v and entity v d The common connection between them is recorded as the layer corresponding association (v d ,θ v ), that is, the corresponding association between entities and attributes; In the attribute layer, the corresponding attribute of each entity v∈V is recorded as Where i represents the type of attribute, and by processing the original information network, the entity layer, the attribute layer and the layer corresponding association are obtained, and then a hierarchical information network is constructed by the entity layer and the attribute layer and the layer corresponding association; The process of step S102, attribute characterization, is as follows: Extract an attribute network G′ from the attribute layer in the hierarchical information network Θ =(Θ,E Θ , R Θ ) to reflect the co-occurrence association between attributes, the co-occurrence of two attributes corresponding to the same entity in the entity layer is recorded as an edge in the attribute network, the number of co-occurrences is used as the weight ω of the association between the attributes, and the edge is recorded as where r′∈R Θ Depends on two properties and Type, A pre-trained word vector is attached to each attribute to reflect the semantic similarity between attributes. Furthermore, the pre-trained word vector is used as the vectorization of the attribute and the pre-trained word vector is recorded as Then, Denoted as a semantic attribute network, where is a word vector set; Using network representation learning to learn semantic attribute network G Θ The embedding vector of the node in the middle, based on the embedding vector, each attribute θ v Denoted as vector The process of step S103, entity characterization, is as follows: Extract the entity network G from the entity layer of the layered information network V =(V, E, R), based on the attribute embedding vector described in step S102, each attribute θ v Convert to the corresponding vector Then, the attribute vector of each entity is connected by combining the entity network and the layer correspondence between the attribute and entity layers Furthermore, the entity network with attributes is recorded as in is the attribute θ v The corresponding vector; In the entity network G with attributes V Apply the network representation learning algorithm to obtain the entity representation corresponding to each entity v in All associations in the data are preserved; The process of step S104, fraud prediction, is as follows: First, two automatic embedding features are designed to reflect the association between loan applications. The automatic embedding features are: observing the similarity between the application identifier and each connected entity from a micro perspective, and observing the similarity between the application identifier and other indirectly connected identifiers from a macro perspective; Introduce entity v into the entity network with attributes t and its first-order neighbor entity v d As a micro-feature, the similarity of entity v t The corresponding embedding vector and entity v d The corresponding embedding vector Calculating Similarity Further, using the entity network v t The second-order neighbor entity v′ t As a macro feature, given the entity v t The corresponding embedding vector and entity v′ t The corresponding embedding vector Calculating Similarity Design manual statistical features based on prior knowledge; Then, the automatic embedding features and manual statistical features are fed into a commonly used classifier to fit the training samples to obtain a fraud prediction model. For the input test sample, after the feature generation, it is input into the fraud prediction model for judgment to obtain the probability that the test sample is fraudulent.

2. The method for predicting online loan fraud based on multi-stage data representation as claimed in claim 1, characterized in that: When selecting fields for the original online lending data, fields with a missing rate lower than 50% are selected as available fields.

3. The online loan fraud prediction method based on multi-stage data representation as claimed in claim 1, characterized in that: Dividing the nodes in the original information network into entities and attributes according to the principles of uniqueness and indivisibility; The uniqueness means that the selected field uniquely represents the same object among multiple fields describing an object; The indivisibility means that an entity does not contain potential information, that is, if a unique value is selected as an entity, then the entity has no overlapping meaning with other entities; The fields that meet the uniqueness and indivisibility are divided into entities, and the set V of entities is recorded as the entity layer in the hierarchical information network; the remaining fields after the entity is selected are divided into attributes, and all attributes θ v The set Θ of is recorded as the attribute layer in the hierarchical information network.

4. The method for predicting online loan fraud based on multi-stage data representation as claimed in claim 1, characterized in that: The missing similarity in the micro-features is replaced by a constant term, and only the first s features with the smallest similarity are retained as macro-features.

5. The method for predicting online loan fraud based on multi-stage data representation as claimed in claim 1, characterized in that: Calculating the similarity using Euclidean distance; For entity v t and v d The embedding vector of , where D is the dimension, for The value of the i-th dimension of for The i-th dimension value of is calculated using the following Euclidean distance formula: For the embedding vector and for The similarity is obtained by using the Euclidean distance calculation formula.

6. The method for predicting online loan fraud based on multi-stage data representation as claimed in claim 1, characterized in that: Attribute network G with semantics Θ and the entity network G with attributes described in S103 V has a similar form, where the attribute set Θ and the entity set V play the same role in the network representation learning algorithm; or, where the pre-trained word vector set G Θ and the set of embedding vectors of attributes Plays the same role in network representation learning algorithms; Therefore, the same network representation learning algorithm is used in both the attribute representation and entity representation stages to integrate different types of information into the corresponding embedding vectors.

Images